arXiv CSopen access

Mobile Imaging Solutions for Medical Diagnosis: Trends and Applications

Mobile Imaging Solutions for Medical Diagnosis: Trends and Applications

arXiv:2609.24814v1 [cs.CV] 21 Sep 2026

Syed Muhammad Ibne Zulfiker1 [email protected]

Tanzima Hashem2 [email protected]

Fariha Tabassum Islam3 [email protected] Md Sultanul Arifin4∗ [email protected] Faria Huq6,a [email protected]

Khandker Aftarul Islam5,a [email protected]

Nishat Anjum Bristy6,a [email protected]

Priyeta Saha7,a [email protected]

Syeda Nahida Akter6,8,a [email protected]

Arpita Saha9,a [email protected]

1

University of Toronto, Toronto, ON, Canada Bangladesh University of Engineering and Technology, Dhaka, Bangladesh 3 University of Wisconsin-Madison, Madison, WI, USA 4 Virginia Tech, Blacksburg, VA, USA 5 Looop Inc., Tokyo, Japan 6 Carnegie Mellon University, Pittsburgh, PA, USA 7 Miro, Amsterdam, Netherlands 8 NVIDIA, Santa Clara, CA, USA 9 Purdue University, West Lafayette, IN, USA 2

Abstract INTRODUCTION: Advances in processing power, sophisticated camera technologies, and mobile image analysis techniques have made smartphones and other mobile devices, such as laptops, suitable for medical diagnosis and healthcare applications. Researchers have developed novel, low-cost solutions for the early detection and monitoring of various health issues—including eye and ENT diseases, malnutrition, heart rate variability, skin and oral conditions, and injuries—using images captured using non-medical devices such as smartphones and webcams. OBJECTIVES: This survey examines existing research that analyzes mobile images to develop medical diagnosis tools, emphasizing their role in enabling lowcost and accessible healthcare solutions. METHODS: We perform a comparative analysis of state-of-the-art solutions, and their advantages and limitations across different healthcare application categories for medical diagnosis. RESULTS: The desired characteristics for medical diagnosis tools using images captured by smartphones or other mobile devices are identified, highlighting areas of achievement and those requiring further research. Application-specific and ∗

Corresponding author.

common challenges are discussed, along with future research directions. CONCLUSION: This comprehensive study provides insights to advance mobile healthcare solutions, enabling low-cost disease diagnosis and monitoring, particularly for underserved populations in remote and resource-constrained areas. Keywords: Survey, Mobile Image Analysis, Smartphone, Healthcare, Diagnosis, Telemedicine

1

Introduction

Technological advancements in personal and mobile devices, particularly the availability of powerful integrated high-resolution cameras in smartphones and laptops, are reshaping healthcare delivery, making pervasive medical solutions more accessible. Despite progress in the healthcare sector, many individuals remain deprived of these benefits due to financial constraints, inadequate facilities, and a shortage of skilled workers. Mobile imaging addresses this gap by bringing essential diagnostic services within reach for people across diverse age groups and socio-economic backgrounds. While it cannot replace comprehensive clinical testing, it significantly improves access to initial screenings. In this survey, we provide a comprehensive review of existing research on mobile imaging solutions for medical diagnosis applications.

Figure 1: Mobile image analysis in healthcare applications Images (including videos) captured with smartphones or laptop cameras can reveal crucial diagnostic cues from body parts such as the skin, eyes, or mouth, often unnoticed by the human eye. Unlike traditional imaging techniques like CT scans or MRIs, these devices provide an accessible means to detect a variety of health conditions, such as skin diseases, eye conditions, and early signs of malnutrition. By applying deep learning, machine learning, and other image analysis methods, meaningful features can be extracted from these images, enabling early detection of potentially lifethreatening diseases. Our study highlights the techniques employed for early disease detection using image captures with non-medical devices, evaluates their strengths and limitations, and identifies key challenges that remain unaddressed. Although several surveys on smartphone-based healthcare applications exist, their scope differs from ours. Most prior studies focus on specific diseases or applications without addressing the broader challenges of adopting mobile imaging solutions for medical diagnosis. For example, [1] reviews smartphone applications for managing depression, while [2] focuses on heart rate estimation, wound assessment and monitoring, and preliminary skin cancer detection. Similarly, [3] explores tools for monitoring physiological parameters. In contrast, [4] provides a broader perspective, with relatively less focus on disease diagnosis among its 15 categories of healthcare applications, which include drug references, clinical communication, Hospital Information Systems (HIS), and medical education. Recent advancements, however, have expanded the scope of smartphone applications by integrating technologies such as molecular analysis, biosensors, mathematical algorithms, microfabrication, 3D printing, and microfluidics. For example, [5] demonstrates how these innovations enable 2

smartphone applications in hematology, digital pathology, and rapid infectious disease diagnostics. Moreover, [6] examines the potential of smartphones in biomedical imaging, further highlighting their versatility in healthcare. Our contributions in this survey paper are as follows: • We provide a comprehensive review of medical diagnosis solutions using mobile images. • We present existing research in different categories and summarize the used techniques and pros-cons. Our comparative analysis facilitates the understanding of existing works in a common structure. • We identify both application-specific and common challenges of mobile imaging solutions for medical diagnosis and examine how these challenges have been addressed in the literature. • Finally, we provide direction for future researchers and draw attention to some important aspects that remain unexplored or less explored. This paper is organized as follows. Section 2 categorizes existing works based on application types. Section 3 provides a comparative review of application-specific research. Section 4 discusses common features and performs a general comparative analysis. Section 5 identifies challenges and future directions, and Section 6 concludes the survey.

Figure 2: Taxonomy graph for included diseases categorized by application area

2

Taxonomies

In Fig. 1, we present a general mobile image analysis workflow for smartphone-based medical diagnosis applications. The process begins with capturing multimedia input of the relevant body part. Next, the input undergoes preprocessing, such as segmenting the region of interest (ROI), followed by feature extraction. These extracted features are then passed through a trained model to detect potential health issues. The model itself is trained using the extracted features from images of a specific body part. Since training a model can be computationally intensive, many researchers leverage cloud computing, where training data is uploaded to a central server equipped with powerful hardware capable of efficiently training large models. The trained model weights are then sent back to the device for use in healthcare applications aimed at early disease diagnosis. Alternatively, edge computing offers a decentralized approach, where both model training and disease inference are performed directly on the device, enabling greater autonomy and reducing dependency on external servers. 3

In Figure 2, we categorize the diseases for which early diagnostic and monitoring tools have been developed, and present a taxonomy graph. In Table 6, for each category, we mention the information regarding what they are, their consequences, symptoms and standard diagnosis measures. We observe that most of the standard methods of diagnosis or measurement are expensive and require a person to visit a medical center. Most of the diagnoses are dependent on the expertise of a healthcare professional. We also observe that the diseases cause visible symptoms. Researchers have attempted to capture these symptoms using mobile image analysis to design low-cost, accessible diagnosis methods. Review Methodology. For selecting and narrowing down relevant publications for each categorized disease, we followed a systematic approach. Our representative search string in Google Scholar includes the following keywords: Smartphone based + {disease name} + (detection OR screening OR diagnosis OR evaluation) + {disease specific keywords} For example, for the disease "Retinopathy of Prematurity (ROP)", our search string was "smartphone based (retinopathy of prematurity OR rop) (diagnosis OR screening) (fundus imaging)". The search results were then further assessed for eligibility on the grounds of i) in or out of scope for this review, ii) level of impact, iii) full text availability, and iv) selection of a single article from the same studies. Further expert judgement was used to focus on relevant publications that used state-of-the-art and novel approaches to solve difficult problems.

3

Comparative Analysis - Application Specific

In this section, we compare the existing research works for each category mentioned in Table 6. We thoroughly evaluate their image analysis pipeline: input, preprocessing steps, features, the classifiers employed, dataset, the output of the analysis, and performance. We identify the desired characteristics for a smart healthcare application and based on them, we check the pros and cons of the existing solutions. Some of these criteria are same for all categories; for example, if a solution does not require internet or skilled workers then it is suitable for people from destitute areas. On the other hand, some desired criteria are application specific; for example, if a diagnostic tool for eye diseases requires eye dilation then users will be not comfortable to use it. We summarize our findings in a comparison table for the health application sectors, where a considerable number of prominent studies are currently available. We observe that there is rarely a solution that fulfills all desired characteristics. Thus, our analysis and findings identify the limitations of the existing solutions and require further research. 3.1 3.1.1

Eye Diseases Diabetic Retinopathy

In the last few years, there have been some significant works [7] on smartphone-based image analysis in diabetic retinopathy (DR) to make eye exams more affordable. These works have captured diagnosable retinal images or videos [8] using smartphone camera with the help of low-cost extensions, e.g. Remidio fundus on phone (FOP), Paxos Scope adapter, RetinaScope etc. These images were graded manually (ground truth) and in automated ways for DR screening. The automatic grading of mobile images has shown high (more than 80%) specificity (SP) and sensitivity (SN) in most of these works. Among the automated techniques used, two are commercially available softwares for DR screening that use deep learning - (i) EyeArtTM software [9] and (ii) Medios AI [10]. [11] assesses the diagnostic accuracy of image interpretations captured using EyeFundusScope. [12] presents a cost-effective and lightweight mobile phone-based system for the early detection of diabetic retinopathy using artificial neural networks (ANN). Table 1 shows the pros and cons of existing works. 3.1.2

Retinopathy of Prematurity

Some works [15, 16, 17, 18, 19] in the last decade have validated that retinopathy of prematurity (ROP) in infants can be detected with reasonable accuracy using smartphone captured fundus images [17] and videos [15, 16, 18, 19]. [20] compared the image quality of different smartphone-based 4

Table 1: Pros and cons of representative literature in diabetic retinopathy Specifications Requires internet connection for high performance computing

[7] ✓

[13] ✗

[10] ✗

[8] N/S

[14] ✓

Requires skilled users

✓

✗

✓

✗

✗

Requires additional equipment

✓

✓

✓

✓

✓

Considers dark and noisy environ- ✗ ment

✗

✗

✗

✗

Considers movement artifacts

✗

✗

✗

✗

✗

Model integrated into a mobile application

N/A

✓

✗

✓

N/A

Performance validation w.r.t gold standard dilated eye test

✗

✗

✗

✗

✓

Requires pharmacological eye dilation

✓

✗

✓

N/S

✓

portable retinal imaging systems and their automatic DR detection accuracy using a deep learning framework. 3.1.3

Cataract

The use of eye images captured with smartphones has also been explored in automatic cataract detection [21, 22, 23, 24, 25, 26, 27]. Lee et al. employed an optimized classifier based on feature elimination combined with a neural network for cataract classification [21]. Similarly, Kaur et al. utilized a neural network for cataract detection [23]. The studies in [25, 24] applied K-nearest neighbors for this purpose. Lau et al. proposed using the red-eye effect as a detection method [22]; however, this approach remains underdeveloped. Only [27] utilized a support vector machine for cataract detection after processing mobile images. 3.1.4

Glaucoma

Glaucoma screening from mobile images has been studied in [28, 29, 30, 31, 32, 33, 34] and concluded that mobile images have the potential to be used in glaucoma screening. Among them, only [32] automated the screening through an interpretable pipeline that uses CNN. 3.2 3.2.1

Malnutrition Body Weight

Body Mass Index (BMI) is an widely used indicator of body weight. Some researchers have worked on BMI estimation from images. Jiang et al. [35] collected 2950 frontal body photos, BMI measurements, and other pertinent data for a visual-body-to-BMI dataset. They extracted anthropometric features from images using a convolutional pose machine (CPM) and estimated BMI using support vector regression (SVR) and Gaussian processing regression (GPR). They show that body image to BMI prediction is feasible and usually more accurate than facial images. Jiang et al. [36] collected an RGB-D video dataset from 163 individuals using the Kinect sensor. They preprocessed the images using Skeleton joints mapping and estimated body volume from derived 3D data. To predict weight from images of dressed people, they developed two clothes models to exclude the effect of clothes on body volume computation. In contrast, Wen and Guo [37] targeted the correlation between facial features and BMI and used 14,500 facial images to conduct their experiments. They compute seven facial features using the active shape model (ASM) and ran SVR on those to produce BMI values. Affuso et al. [38] measured obesity using body fat percentage. They took three images of 323 — front, back, and side profiles — and used the front curve and side curve to construct body form features. Then, they cluster participants using K-means and Calinski-Harabasz criterion. Finally, 5

they trained their SVR model to predict body fat percentage. [39] attempted to identify malnutrition and predict BMI using real-time facial image processing. [40] assessed the validity of exisiting image-assisted mobile nutrition apps. 3.2.2

Hemoglobin Analysis

Many alternatives to blood hemoglobin level testing for anemia diagnosis have been proposed in recent years [41, 42]. HemaApp, a smartphone app by Wang et al. [43], measures blood hemoglobin concentration noninvasively. They attached IR light sources of different wavelengths in hardware. The app records fingerlight reflection videos for each light source. It turns these videos into RGB time-series waveforms, identifies features, and trains SVM regression models based on blood test hemoglobin levels for each wavelength. HemaApp is effective in anemia classification (with 85.7% sensitivity and 76.5% specificity). This work was extended to HemaApp V.2, by using the smartphone’s back camera and white flash LED [44], and achieved an RMSE of 1.27. Mannino et al. went one step further and developed a standalone app to measure blood hemoglobin levels without extra equipment or smartphone attachment [45]. They used this app to prompt their subjects to take photos of their fingernails and select the ROI. Fingernail and skin color data were extracted from these ROIs and fed into a robust multi-linear regression algorithm with bi-square weighting. This system diagnosed anemia with 92% sensitivity and 76% specificity. Hasan et al. [46] recorded 10-second fingertip videos, generated color maps of pixel color (RGB) intensity for each frame, and then segmented the images into blocks to capture pixel value changes. They trained their ANN model with block mean pixel values. The model predicted hemoglobin levels with 94% sensitivity and 96% specificity on their test dataset. Hasan et al. [47] proposed another approach that converted RGB pixel intensities into HSV, Lab, and gray color spaces. They predicted hemoglobin levels from color space values using Partial Least Squares (PLS). [48] proposed a system which can efficiently access blood haemoglobin levels non-invasively by observing color changes in nail-bed. [49] introduced a novel smartphone-based system for non-invasive hemoglobin measurement via eyelid imaging, utilizing deep learning algorithms such as EGE-Unet and DHA (C3AE) models. This system demonstrated greater accuracy and efficiency compared to traditional methods and manual assessments. 3.3

Heart Rate Variability

There are two standard methods for measuring heart rates using smartphone applications: (i) from fingertips [50, 51], which has evolved from a contact-based method to a non-contact based method [52] (ii) from facial images and videos, which is essentially a non-contact method. Following, we analyze existing studies on these two methods. Table 2 discuss the pros and cons of those works. 3.3.1

Fingertip-based Heart Rate Measurement

Sukaphat et al. measured heart rate using the image of a single fingertip [50]. This study employed the mean value of RGB color channel histograms to determine heart rate frequency. This approach matched a digital pressure monitor by 0.57±% when tested on 10 people. Conversely, Zaman et al. used successive fingertip images to calculate heart rate [51] using two methods. In the first method, the authors applied edge detection and smoothing, and then, measured the change of the fingertip curve lines with time, which can be converted into time and frequency domain graphs to detect heart rate. In the second method, the authors used the changes in the image intensity of the red channel of the successive fingertip images to compute heart rate. 3.3.2

Facial Images based Heart Rate Measurement

There are currently two remote methods for measuring heart rate (HR) from face videos: color intensity-based and motion-based. Photoplethysmography (PPG) signals captured by a camera are utilized in light-intensity based methods that measure variations in blood volume by detecting changes in light reflection or transmission during the cardiovascular pulse cycle. This idea was first introduced in 2008 by Verkruysse et al. [61] and later got adopted by [53, 54, 55, 57, 56]. The heartbeat signal is extracted using blind source separation (BSS) on facial color changes. The ROIs are selected from the videos, although the relative importance of the ROIs is still debatable. Independent Component Analysis (ICA) is the most common BSS method used to derive the face BVP signal. Initially, Poh et al. [53] and Kwon et al. [55] used the second component of the ICA (i.e., the green channel) and attained RMS error less than 5 bpm compared to Flexcomp finger BVP sensor 6

Table 2: Pros and cons of representative literature in heart rate variability Specifications Requires internet connection for high performance computing

[50] ✗

[51] ✓

[53] N/S

[54] N/S

[55] ✗

[56] N/S

[57] N/S

[58] N/S

[59] N/S

[60] N/S

Requires skilled users

✗

✗

✗

✗

✗

✗

✗

✗

✗

✗

Requires addi- ✗ tional equipment

✓

✗

✗

✗

✗

✗

✗

✗

✗

Considers ✗ dark and noisy environment

✗

✗

✗

✗

✓

✗

✓

✓

✓

Considers movement artifacts

N/A

N/A

✓

✗

✗

✓

✓

✓

✗

✗

Model integrated into a mobile application

✓

✓

✗

✗

✓

✗

✗

✗

✗

✗

✗ Support mobile devices/ cameras with variant configuration

✗

✗

N/S

✓

✗

✗

✗

✗

✗

HR estimation ✗ for multiple person at a time

✗

✓

✗

✗

N/S

N/S

✗

N/S

N/S

Performance ✓ validation w.r.t ground truth

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✓

✗

✗

Requires RGB ✓ color images/ videos

✓

✓

✓

✓

✓

✓

✓

✗

✗

Requires skin to be visible

7

and average error rates of 1.04% compared to ECG, respectively. Later, other studies [54, 57, 56] have analyzed all ICA components and improved performance. Though ICA and Principal Component Analysis (PCA) work experimentally well, they consider an observed signal to be a linear combination of multiple sources, which conflicts with the Beer-Lambert law that states that the intensity of light reflected through facial tissue is a nonlinear function of distance [62, 63]. Some other dimensionality reduction algorithms used to extract face BVP signals are linear discriminant analysis, locally linear embedding, and manifold learning methods such as Isomap, Laplacian Eigenmap (LE) [64]. Wei et al. [63] investigated nine such algorithms on RGB color channels and found that Laplacian Eigenmap extracted BVP information best. In contrast, HR detection methods based on motion originated from ballistocardiogram. Ballistocardiography is a technique used to record the repetitive motions of the human body in the form of ballistocardiogram signals. Balakrishnan et al. [59] were the were the first to propose analyzing head motions during remote HR detection from face videos. They used PCA to project the pulse generated from head motion to 1D signals. This method yielded an error rate of 2.2% when compared to ECG outputs, indicating that methods analyzing facial video provide superior performance. Shan et al. [60] chose a single feature point from the head (instead of considering all points in the head [59]) and demonstrated that ICA is more consistent and accurate than PCA.. 3.4

Skin Diseases

3.4.1

Skin cancer

Hameed et al. compiled studies up to 2016 on skin cancer using image analysis techniques [65]. In those studies, images were preprocessed by scaling, filtering, cropping, and hair removal. Colorbased, region-based, and threshold-based segmentation approaches were employed. The most popular classifier was SVM, followed by ANN and then KNN. Later, Brinker et al. summarized CNNbased studies for diagnosing skin cancers through 2018 [66]. Some of those studies employed CNN as a feature extractor, but the majority used transfer learning with CNN for end-to-end learning. Esteva et al. trained a deep CNN to classify images with similar accuracy to a dermatologist [67]. Later, Brinker et. al employed a pre-trained ResNet50 model that outperform dermatologists [68]. CNN was also used to diagnose and categorize skin cancer in several recent research [69, 70, 71, 72]. Additionally, Kumar et al. used fuzzy C-means clustering for segmentation and trained ANN using a meta-heuristic algorithm, differential evolution, to successfully detect skin cancer [73]. All these studies were desktop-based, however, SkinScan©, a light-weight portable library for the automatic detection of melanoma suitable for smartphones and handheld devices, was developed in 2011, [74] and Kalwa et al. recently developed an end-to-end smartphone application for identifying melanoma with 88% accuracy [75]. Table 3 highlights the pros and cons of the existing research works. Note that, most of these works have not evaluated the performance of their skin cancer identifying models for smartphone captured images. Table 3: Pros and cons of representative literature in skin cancer Specifications

[76]

[69]

[75]

[68]

Requires skilled users

✗

✗

✗

✗

Requires additional equipment

✗

✗

✓

✗

✗

N/S

N/S

✗

✗

✗

✓

✓

✗

N/S

✗

✗

✗

✗

✓

✓

✓

Considers dark and noisy environ- N/S ment Model integrated into a mobile application

✓

Supports devices with variant con- N/S figurations Performance validation w.r.t ground truth

✓

8

[73]

3.4.2

Inflammatory Skin Diseases

Several studies [77, 78, 79] examined the diagnosis procedure using a mobile application. Gupta et al. utilized Gaussian Mixture Models due to their greater computational efficiency and resourcefriendliness, making them suitable for smartphones with resource constraints, unlike heavier deep learning models [77]. Alamdari et al. studied low-resolution input images fed to the machine learning models as images from smartphones are expected to be low-resolution [78]. Yasir et al. surveyed the doctors in a developing country where skin diseases are widely prevalent and developed a system to detect the nine most common diseases by studying the benefits and usability of such an application [79]. Some researchers developed computer-based systems that use digital skin images from different websites as input, which can be extended easily to mobile applications using cross-platform developments [80, 81, 82]. Many of these works performed heavy preprocessing using various image processing techniques to reduce noise and sensitivity to lighting conditions and finally to segment the affected areas from the image before feeding them into the model. Mostly two types of features, color and texture, were extracted before the classification task indicating that these features aid the model in giving accurate predictions. Although there was no significant usage of statistical models as classifiers, lightweight machine learning models were mostly used with SVM gaining popularity among them. In a few studies such as [83] and [79], external features such as elevation were input to the system by the user along with the skin image, but the majority of the works in this domain used only skin image as input. In some studies [84, 78], the system uses metrics like EASI and other scoring algorithms to determine the severity of the illness, which might assist users to comprehend the gravity of their condition and decide what steps to take next. Table 4 summarizes the pros and cons of existing works related to inflammatory skin diseases. Table 4: Pros and cons of representative literature in inflammatory skin diseases Specifications [83] Requires internet connection for high perfor- ✗ mance computing

[77] ✗

[78] ✗

[80] ✗

[84] ✗

[85] ✗

[86] N/S

Requires skilled users to operate the system

✓

✗

✗

✗

✗

✗

✗

Requires additional equipment

✗

✗

✗

✗

✗

✗

✗

Requires additional information input by a ✗ trained professional

✗

✗

✗

✗

✗

✓

✗

✗

✗

✗

✗

✓

✗

Accounts for other objects present in the im- ✓ age except skin

✗

✓

✗

✓

✗

Able to identify multiple diseases or occur- ✓ rence of a disease from a single image

N/S

✓

N/S

✓

✗

N/S

Model integrated into a mobile application or ✗ a computer system

✗

✗

✓

✗

✗

✓

Device specific performance validation

✗

✗

✓

N/S

✗

✗

Considers dark and noisy environment

3.4.3

N/S

Other

Chan et. al made a comprehensive compilation of machine learning applications in dermatology [87]. In most cases, the models used for classifying dermatological diseases (e.g. melanoma, nonmelanoma skin cancer) were as accurate (based on metrics like sensitivity, specificity, AUC) as the diagnoses done by board-certified dermatologists. However, due to the lack of external validation, these models often might not yield expected results for other real-life datasets. The review also highlighted the fact that existing mobile applications for classifying skin diseases are still not accurate enough to be clinically approved although they can be highly useful in solidifying the diagnosis process for dermatologists. In addition, some crucial limitations of machine learning (e.g. black-box nature of ML algorithms, dependency on input quality, algorithmic bias, lack of inclusivity, clinical regulations) were discussed in [87]. 9

3.5

Injury Detection

3.5.1

Skin Burns

Pande et al. proposed a method for categorizing burn levels and recommending first-aid treatments based on these levels [88]. Chauhan et al. utilized ResNet50 to classify burnt body parts, aiming to automate the early assessment of the total body area affected by burns [89]. Abubakar et al. conducted pioneering work on classifying skin-burn images across different racial groups, specifically focusing on Caucasian and African skin types [90]. More recently, an increasing number of studies have employed CNNs and transfer learning techniques (e.g., pretrained ResNet, VGG16) to distinguish between burnt and injured skin [91] or to assess burn depth [92, 93]. Table 5 summarizes the advantages and limitations of the various studies on skin burn detection and classification. Table 5: Pros and cons of representative literature in skin burn Specifications Requires internet connection for high performance computing

[88] ✓

[94] ✓

[90] ✓

[93] ✓

[95] ✓

Requires skilled users

✗

✗

✗

✗

✗

Requires additional equipment

✗

✗

✗

✗

✗

N/S

N/S

N/S

N/S

✗

✗

✗

✗

✗

Supports devices with variant con- ✗ figurations

✗

✗

✗

✗

✓

✓

✓

✓

✓

Considers dark and noisy environ- N/S ment Model integrated into a mobile application

Performance Validation w.r.t ground truth

3.5.2

Wound Detection

A variety of wound assessment systems have been developed to aid diabetic patients and monitor healing. In [96, 97], an accelerated mean-shift algorithm was used for wound segmentation, followed by a connected region detection method to determine wound boundaries. The healing state was then measured using red-yellow-black color rating models based on trend analysis of patient time records. Another system for diabetic wound assessment employed the K-means shift algorithm [98]. More recently, Shenoy et al. developed Deepwound, a multi-label CNN ensemble capable of classifying wound images using only pixel data and labels [99]. Kaile et al. introduced a smartphonebased near-infrared imaging tool to quantify tissue oxygenation, a critical factor in wound healing. Changes in oxygen levels were shown to correlate with healing status [100]. Furthermore, Lau et al. presented a real-time pressure injury wound detection and staging classification system using YOLOv4 [101]. This application highlights the growing potential of real-time machine learning models in clinical wound assessment. 3.5.3

Brain Injury Detection

Mariakakis et al. developed a smartphone-based system for detecting concussions by assessing pupillary light reflex [102]. Using a smartphone flashlight to stimulate the eyes and recording the response via video, they employed a fully convolutional network (FCN) to track pupil diameter changes over time. This analysis allowed extraction of clinically relevant metrics, including constriction amplitude, percentage, and velocity. To minimize bias, the dataset included various iris colors, and conservative cropping excluded skin areas. The app, tested on individuals with mild traumatic brain injury (mTBI), offered a low-cost and portable pupillometer. 10

Additionally, [103] introduced a smartphone-enabled optofluidic platform for the rapid detection of brain-derived exosomes, offering a faster and non-invasive method to identify potential biomarkers for concussion recovery compared to traditional techniques. 3.6

Oral Diseases

3.6.1

Oral Cancer

Several techniques have been employed in the literature to detect oral cancer from intraoral and whole-cavity imaging. Previously, Anuradha et al. proposed a system to detect benign or malignant tumors using the SVM classifier [104]. Recently, smartphone-based probes have shown remarkable results in detecting oral cancer using CNN [105, 106, 107]. These studies demonstrated sensitivity and specificity ranging from 81.25% to 94.94%, compared to specialists. "[108] proposed a new image-capturing method for consistent lesion positioning and used HRNet for oral cancer classification. [109] developed a novel automatic oral cancer detection algorithm to identify and differentiate premalignant lesions from buccal cavity images, aiding in the early detection of oral cancer. [110] used a cascaded CNN to detect oral cavity squamous cell carcinoma (OCSCC) from photographic images. 3.6.2

Tongue Diagnosis

The review paper by Jung et al. [111] summarizes studies up to 2012, where most research relied on data collected using external hardware for tongue diagnosis. Since then, the required devices have evolved from relying on external equipment to using only smartphones. Zhang et al. [112] applied SVM and KNN algorithms to confirm the link between human health and tongue color. They categorized tongue images into healthy and diseased with 91.99% accuracy, and detected 11 diseases with 70% accuracy from a dataset of 1,045 tongue images (143 healthy and 902 diseased). In this study, tongue images were assigned to one of 12 color gamuts for classification. Another study [113] introduced or capturing tongue images, utilizing an optimized Canny algorithm to detect tongue outlines, followed by statistical analysis for automated diagnosis. Ryu et al. [114] developed a tongue color calibration method to correct smartphone-captured tongue images by adjusting the white balance of teeth color, compensating for the effects of surrounding light exposure. In contrast, Hu et al. [115, 116, 117] proposed an SVM-based approach for estimating lighting conditions, predicting a color correction matrix based on the color difference between images taken with and without a flashlight. Kanawong et al. [118] conducted computer-aided tongue image analysis based on traditional chinese medicine. Most recently, Huang et al. [119] developed an automated tongue diagnosis system using Mask R-CNN for tongue-image segmentation and ResNeXt for tongue-coating color classification. Their mobile app implementation achieved significant improvements in accuracy for non-invasive disease assessment in traditional Chinese medicine. 3.6.3

Dental Diseases

A few notable works, such as [120, 121], have focused on caries detection using smartphone images. Zhang et al. [121] proposed an automated smartphone application for detecting early childhood caries, which was evaluated by trained dentists but still requires improvements to match their performance levels. They employed a Single Shot Multibox Detector with a MobileNet-v2 backbone as the classifier, which is lightweight and well-suited for resource-constrained smartphones, enabling faster, real-time inference without requiring internet connectivity. On the other hand, Choi et al. [120] addressed lighting effects in images, enhanced image quality, and applied adaptive hue thresholding to extract the teeth region. They also used erosion and dilation techniques to improve image accuracy. Their Android application took user input (images) and performed all computations on a cloud server, requiring an internet connection. In 2019, Askarian et al. [122] studied the use of smartphones for detecting periodontal diseases. They collected 30 images (15 diseased and 15 healthy) using an iPhone X rear camera with an addon gadget to reduce refraction effects and ambient light interference. The images were preprocessed using color tuning and correction to reduce noise and extract the gum region (ROI). A gum color 11

gamut was designed to represent all possible gum surface colors, and features were extracted from the CEILAB color space. SVM was used as the classifier, achieving an average accuracy of 94.3%, specificity of 93%, and sensitivity of 92.6%. Finally, You et al. [123] introduced a deep learning-based AI model for detecting dental plaque on primary teeth. Users uploaded smartphone-captured images, and the model demonstrated clinically acceptable diagnostic accuracy, achieving a mean intersection-over-union (MIoU) of 0.736, comparable to the performance of an experienced pediatric dentist. 3.6.4

Miscellaneous Oral Diseases

To enable users to regularly inspect their oral health independently, Liang et al. developed an interactive mobile app named OralCam [124]. They employed a Multi-Task Learning approach with a single deep CNN model to extract features from input images capturing five different oral conditions: periodontal diseases, caries, soft deposit, dental calculus, and dental discoloration. In the second stage of their algorithm, another deep CNN model was used to locate bounding boxes for each finding. To train their model, the researchers prepared an in-house dataset of 3,182 oral cavity images annotated by dental experts. The dataset also included information about users’ living habits, history of pain, and bleeding. In addition to traditional evaluation metrics like accuracy and sensitivity, they incorporated feedback from end users and board-certified dentists to improve the app’s robustness. This work demonstrates significant promise in creating accessible and functional mobile health solutions. 3.7 3.7.1

ENT Ear Infection Detection

Otitis Media (middle-ear infection) can be detected using a smartphone with a plug-in otoscope device. Huang et al. [125] captured images using an otoscope and applied an active contour algorithm to segment the region of interest (ROI). Different Otitis Media cases were then classified using a Depth-First Search algorithm. The images were transmitted to a server for analysis. Myburgha et al. [126] used a neural network to classify normal and Otitis Media cases. They employed a portable digital otoscope to capture images, which were saved and displayed on a smartphone before being processed on a dedicated server. In a more recent development, Chen et al. [127] designed a smartphone-adapted deep learning model for detecting and diagnosing middle-ear diseases. They leveraged transfer learning and lightweight CNNs such as MobileNet, making the system highly efficient for use on mobile devices. 3.7.2

Throat Diseases

The use of smartphones for detecting Streptococcal Pharyngitis (strep throat) has been explored in [128, 129]. Askarian et al. compiled high-resolution throat images (4032×3024 pixels) with metadata like age and sex, while Yoo et al. used low-resolution images (256×256 pixels) without metadata. Yoo et al. addressed the issue of limited datasets by collecting 131 pharyngitis and 208 normal images, classified by clinicians, and augmented the data using CycleGAN to generate paired normal and pharyngitis cases. Both studies used balanced training datasets: 20 strep/20 healthy images in [128] and 1500 strep/1500 healthy images in [129]. For preprocessing, Askarian et al. applied color correction and intensity-based segmentation to extract throat ROIs, leveraging features like redness and swollen tonsils. Yoo et al. skipped explicit preprocessing or feature extraction. Classifier choice also differed: Askarian et al. employed KNN for its lightweight nature, achieving 93.75% accuracy (87.5% sensitivity, 88% specificity), while Yoo et al. used pre-trained CNNs (ResNet50, Inception-v3, MobileNet-v2), with MobileNet-v2 being optimal for smartphones, achieving 95.3% accuracy (97% sensitivity, 94.2% specificity). Additionally, a mobile app for early Tonsillitis detection processes oral cavity images [130]. Ban et al. [131] demonstrated a smartphone-attachable infrared camera for detecting peritonsillar abscesses in tonsillitis patients, showcasing non-invasive thermography as a viable alternative to CT scans. 12

3.8 3.8.1

Others Blood Pressure

Luo et al. [132] proposed a novel contactless method for measuring blood pressure using facial videos captured with a smartphone camera, known as Transdermal Optical Imaging (TOI). TOI separates video frames into RGB bitplanes, extracting hemoglobin-rich signals while discarding melanin-rich ones through an ML-based algorithm. These signals are then merged to generate videos visualizing hemoglobin concentration changes. A total of 1,328 subjects aged 18 or older, with average reference systolic BP between 100–139 mmHg and diastolic BP between 60–89 mmHg, were included in the study. Signal processing identified 17 ROIs, from which 126 facial transdermal blood flow signals were extracted. PCA was applied for dimensionality reduction, and the features were fed into a multi-layer perceptron to predict systolic BP, diastolic BP, and pulse pressure. The model achieved prediction accuracies of 94.81% (systolic BP), 95.71% (diastolic BP), and 95.75% (pulse pressure). Hypertensive and hypotensive subjects were excluded from the study scope. 3.8.2

Jaundice

Jaundice can be detected from smartphone-captured images of a person’s eye by analyzing the sclera (the white portion of the eye) using computer vision techniques. Zhou et al [133] proposed an RGB and HSV color space masking technique for sclera segmentation, while Mariakakis et al. [134] utilized the GrabCut algorithm. Using a dataset of 56 images, Mariakakis et al. achieved 95.7% sensitivity and 97.4% specificity in jaundice detection with a Random Forest classifier. 3.8.3

Parkinson’s Disease (PD)

n video-based analysis of Parkinson’s Disease (PD), the initial step involves selecting frames from the video. Ali et al. [135] utilized a CNN-based classifier to identify the relevant frames. For the final stage, a CNN with a VGG16 architecture [136] is employed to detect important hand-motor movements from the selected frames. The tasks are assessed separately for two-class classification (PD vs. non-PD) or three-class classification (non-PD, PD with medication, PD without medication). The results show that the two-class classification, with approximately 80% accuracy, 80% precision, and 88% recall, is more reliable than the three-class classification, which achieves 60% accuracy, 65% precision, and 60% recall. 3.8.4

Post-traumatic Stress Disorder (PTSD)

Farhana et al. [137] used free-hand sketch images from 160 individuals to diagnose PTSD based on three features: the number of corners, the number of strokes, and the average stroke length. Using only sketching data and the gender of the individuals, their model achieved an accuracy of 82.9–87.8%. When EEG data was combined with sketching data, the model’s accuracy improved significantly to 99.29%. 3.8.5

Mental Health Diseases

Zunino et al. [138] proposed a solution for early autism diagnosis using video sequences, leveraging differences in simple motor actions between pathological and healthy subjects. Similarly, Tian et al. [139] developed a One Glimpse Autism Spectrum Disorder Detection (O-GAD) network, focusing on repetitive behavioral analysis from video streams. 3.8.6

Gait Abnormalities

Aroojis et al. [140] utilized a handheld smartphone with a slow-motion video feature to measure the Edinburgh Visual Gait Score (EVGS) in children with cerebral palsy. Videos of the coronal and sagittal planes were recorded as patients walked, and specific angles were measured by pausing the slow-motion video and using a motion analysis app on the smartphone. Two Pediatric Orthopedic Fellows independently computed the EVGS, achieving complete agreement rates between 61.7% and 92.5%.. 13

3.8.7

Dehydration

In [141], the authors developed a smartphone-based application for dehydration detection using facial image features. The app captures a facial image, calculates dehydration scores for multiple facial landmarks using a Siamese network-based ensemble learning algorithm, and computes a weighted average of these scores to classify whether the individual is dehydrated. Since no public dataset was available, the authors created their own by collecting paired hydrated and dehydrated facial images through the app’s guided interface.

4

Comparative Analysis - General

In this section, we focus on the comparison of some common aspects, such as system architecture, models etc. of the previously discussed works. 4.1

Architecture

(a) Cloud based architecture for smartphonebased image analysis

(b) Offline architecture for smartphone-based image analysis

Figure 3: Architecture for image analysis for medical diagnosis Few studies have thoroughly explored the system architectures for healthcare applications. Two primary types of architectures have been identified: 1. Cloud Based Architecture [7, 14]: In this architecture, models are stored in the cloud, requiring internet connectivity for classification tasks. Smartphones either send the raw input image or extract key features locally before transmitting them to the cloud. This approach allows models to be updated with new training data, improving their generalizability over time. However, reliance on connectivity and data privacy concerns are significant limitations. 2. Offline Architecture [13, 10, 32]: Here, model weights are stored locally on the smartphone, enabling all computations to occur on-device. This architecture is well-suited for low-connectivity environments and ensures data privacy. However, it requires lightweight models to accommodate the computational constraints of smartphones, potentially sacrificing some accuracy. 4.2

Models

Most of studies that focused on detecting healthcare problems from digital images using image processing techniques, statistical and machine learning models were conducted in 2015. Over the past two decades, the popularity of deep learning models, especially CNNs (basic and advanced), has increased significantly. The usage of ANN (Artificial Neural Networks) began in 2002 but reached its peak in 2015. 14

We notice that signal analysis methods are particularly popular in heart rate variability (HRV) analysis due to their faster processing capabilities compared to machine learning and other sophisticated models. These models can also perform real-time HR detection using only a smartphone, with a minimal accuracy gap compared to ground truth values, further enhancing their popularity. SVM (Support Vector Machine) is widely used in domains such as skin cancer detection and inflammatory skin diseases. This is because SVMs can draw fine distinctions between similar skin diseases, with the model’s hyperplane effectively separating the classifications, aiding in accurate predictions. CNNs have been most widely used in diabetic retinopathy, skin cancer, and heart rate variability analysis. Initially, these models were used only for skin cancer and HRV detection, but with the advent of newer, more effective models, their applications have expanded across a broader range of healthcare domains.

5

Challenges and Future Directions • Architecture: Current smartphone applications lack significant deployment of deep learning models, except in limited cases such as diabetic retinopathy, where offline and cloudbased architectures are utilized (see Section 4.1). The gap arises because smartphone processors, while improving, cannot match the computational power required for GPUintensive deep learning tasks. Future exploration could involve designing personalized architectures in which models train locally on user-provided data and offer tailored feedback. Such architectures could revolutionize domains like blood pressure monitoring and diabetes management using smartphones. • Unified performance measure: Healthcare applications within the same domain often lack standardized performance metrics. For example, in heart rate variability analysis, models use varying metrics such as discrepancy, RMSE, or error rates, which complicates crossmethod comparison. Defining and adopting universal metrics for specific healthcare domains can improve consistency and reliability in evaluating models. • Explainability: Deep learning models in healthcare primarily function as black boxes, offering little insight into their decision-making processes. Research to improve the interpretability of the model could lead to the design of more accurate input and increase user trust. Explainability tools such as SHAP [142] or LIME [143] could help identify the features that drive model predictions. • Environmental Robustness: Image-based healthcare applications are often sensitive to environmental factors such as lighting or background noise, which can impact accuracy. Developing methods that eliminate environmental dependencies, such as advanced preprocessing or domain adaptation techniques, is a critical research direction. • User-centered application design: Non-technical users may struggle to input clear and focused images into healthcare apps, leading to suboptimal results. Guided interfaces with features like on-screen alignment guides or stabilization prompts can enhance input quality and usability for non-expert users. • Benchmark dataset: A lack of benchmark datasets across domains like skin disease detection and diabetic retinopathy hampers effective model comparison. Establishing standardized datasets would enable robust evaluation and foster competition, leading to betterperforming models. • Computational constraints: Deep learning models are often computationally expensive, making them unsuitable for low-end smartphones. Research into lightweight architectures using techniques like model pruning, quantization, or knowledge distillation can bridge this gap without sacrificing accuracy. • Ethical and Privacy Concerns: With the increasing use of personal medical data for training and deploying models, ensuring data privacy and ethical use is critical. Research could focus on developing privacy-preserving techniques such as federated learning or differential privacy to ensure that sensitive user data is not exposed. • Real-World Validation: One of the key challenges for mobile healthcare applications is testing them in real-world settings and validating their accuracy. Future work should focus 15

on comparing the results of these diagnostic applications with findings from actual medical diagnoses to ensure their reliability and effectiveness in clinical practice.

6

Conclusion

With rapid advancements in mobile technology, smartphones have seamlessly integrated into our daily lives, enabling transformative innovations across multiple domains, including healthcare. Leveraging these capabilities, researchers have made remarkable progress in developing smartphone-based healthcare applications for early disease detection and monitoring. This survey provides a comprehensive review of the existing work in this domain, focusing on applications that utilize mobile image analysis. We categorized healthcare applications, highlighting their desired characteristics and summarizing the advantages and limitations of current solutions. Our findings identify gaps in the field, such as the need for standardized performance metrics, robust architectures for real-world environments, and improved explainability. These insights aim to inspire further research to address these challenges effectively. Moreover, this study serves as a guide for designing novel, image-based healthcare applications that can cater to emerging disease categories. By tackling the outlined challenges, the potential for creating accessible, low-cost diagnostic tools will significantly improve, particularly for underserved populations in remote and resource-constrained areas. Although the field of smartphone-based image analysis in healthcare is quickly expanding with numerous studies each year, we found that several challenges remain unaddressed or inadequately addressed. Addressing these challenges will make low-cost smart healthcare more accessible to people in remote and underserved areas.

References ¨ [1] Ariane Kerst, J"urgen Zielasek, and Wolfgang Gaebel. Smartphone applications for depression: a systematic literature review and a survey of health care professionals’ attitudes towards their use in clinical practice. European Archives of Psychiatry and Clinical Neuroscience, 270(2):139–152, 2020. [2] Hossein Nejati, Victor Pomponiu, Thanh-Toan Do, Yiren Zhou, Sahar Iravani, and Ngai-Man Cheung. Smartphone and mobile image processing for assisted living: Health-monitoring apps powered by advanced mobile imaging algorithms. IEEE Signal Processing Magazine, 33(4):30–48, 2016. [3] Sandeep Kumar Vashist, E Marion Schneider, and John HT Luong. Commercial smartphonebased devices and smart applications for personalized healthcare monitoring and management. Diagnostics, 4(3):104–128, 2014. [4] Abu Saleh Mohammad Mosa, Illhoi Yoo, and Lincoln Sheets. A systematic review of healthcare applications for smartphones. BMC Medical Informatics and Decision Making, 12(1):1– 31, 2012. [5] Iván Hernández-Neuta, Felix Neumann, J Brightmeyer, T Ba Tis, Narayanan Madaboosi, Q Wei, A Ozcan, and Mats Nilsson. Smartphone-based clinical diagnostics: towards democratization of evidence-based health care. Journal of Internal Medicine, 285(1):19–39, 2019. [6] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60–88, 2017. [7] Ramachandran Rajalakshmi, Radhakrishnan Subashini, Ranjit Mohan Anjana, and Viswanathan Mohan. Automated diabetic retinopathy detection in smartphone-based fundus photography using artificial intelligence. Eye, 32(6):1138–1144, 2018. 16

[8] Simon Mueller, Snezhana Karpova, Maximilian WM Wintergerst, Kaushik Murali, Mahesh P Shanmugam, Robert P Finger, and Thomas Schultz. Automated detection of diabetic retinopathy from smartphone fundus videos. In International Workshop on Ophthalmic Medical Image Analysis, pages 83–92, 2020. [9] Inc. Eyenuk. Eyeart - eyenuk, inc. - artificial intelligence eye screening. https://www. eyenuk.com/en/products/eyeart/, 2010. Accessed: March 24, 2022. [10] Bhavana Sosale, Aravind R Sosale, Hemanth Murthy, Sabyasachi Sengupta, and Muralidhar Naveenam. Medios–an offline, smartphone-based artificial intelligence algorithm for the diagnosis of diabetic retinopathy. Indian Journal of Ophthalmology, 68(2):391, 2020. [11] Sílvia Rêgo, Matilde Monteiro-Soares, Marco Dutra-Medeiros, Filipe Soares, Cláudia Camila Dias, and Francisco Nunes. Implementation and evaluation of a mobile retinal image acquisition system for screening diabetic retinopathy: study protocol. Diabetology, 3(1):1–16, 2022. [12] Nikita Kashyap, Dharmendra Kumar Singh, and Girish Kumar Singh. 4 - mobile phonebased diabetic retinopathy detection system. In Ayman S. El-Baz and Jasjit S. Suri, editors, Diabetes and Fundus OCT, Computer-Assisted Diagnosis, pages 113–131. Elsevier, 2020. [13] Sundaram Natarajan, Astha Jain, Radhika Krishnan, Ashwini Rogye, and Sobha Sivaprasad. Diagnostic accuracy of community-based diabetic retinopathy screening with an offline artificial intelligence system on a smartphone. JAMA Ophthalmology, 137(10):1182–1188, 2019. [14] Tyson N Kim, Michael T Aaberg, Patrick Li, Jose R Davila, Malavika Bhaskaranand, Sandeep Bhat, Chaithanya Ramachandra, Kaushal Solanki, Frankie Myers, Clay Reber, et al. Comparison of automated and expert human grading of diabetic retinopathy using smartphone-based retinal photography. Eye, 35(1):334–342, 2021. [15] Shan-Jiun Lin, Chung-May Yang, Po-Ting Yeh, and Tzyy-Chang Ho. Smartphone fundoscopy for retinopathy of prematurity. Taiwan Journal of Ophthalmology, 4(2):82–85, 2014. [16] T Lekha, S Ramesh, Ashish Sharma, and G Abinaya. Mii retcam assisted smartphone based fundus imaging for retinopathy of prematurity. Indian journal of ophthalmology, 67(6):834, 2019. [17] Maximilian WM Wintergerst, Michael Petrak, Jeany Q Li, Petra P Larsen, Moritz Berger, Frank G Holz, Robert P Finger, and Tim U Krohne. Non-contact smartphone-based fundus imaging compared to conventional fundus imaging: A low-cost alternative for retinopathy of prematurity screening and documentation. Scientific Reports, 9(1):1–8, 2019. [18] Anubhav Goyal, Mahesh Gopalakrishnan, Giridhar Anantharaman, Dhileesh P Chandrashekharan, Thomas Thachil, and Ashish Sharma. Smartphone guided wide-field imaging for retinopathy of prematurity in neonatal intensive care unit–a smart rop (srop) initiative. Indian Journal of Ophthalmology, 67(6):840, 2019. [19] Tapan P Patel, Michael T Aaberg, Yannis M Paulus, Philip Lieu, Vaidehi S Dedania, Cynthia X Qian, Cagri G Besirli, Todd Margolis, Daniel A Fletcher, and Tyson N Kim. Smartphone-based fundus photography for screening of plus-disease retinopathy of prematurity. Graefe’s Archive for Clinical and Experimental Ophthalmology, 257(11):2579–2585, 2019. [20] Mahmut Karakaya and Recep E Hacisoftaoglu. Comparison of smartphone-based retinal imaging systems for diabetic retinopathy detection using deep learning. BMC bioinformatics, 21(Suppl 4):259, 2020. [21] Byung-Kwan Lee, Eun-Hee Jeong, and Yusrina Tifani. An intelligent decision support system for retinal disease diagnosis based on svm using a smartphone. The Journal of Korea Institute of Information, Electronics, and Communication Technology, 8(5):373–383, 2015. [22] Sian Lun Lau, Jan Bond Chan, et al. Mobile cataract screening app using a smartphone. In IEEE Conference on e-Learning, e-Management and e-Services (IC3e), pages 110–115, 2015. 17

[23] Manpreet Kaur, Jaspreet Kaur, and Ravinder Kaur. Low cost cataract detection system using smart phone. In International Conference on Green Computing and Internet of Things (ICGCIoT), pages 1607–1609, 2015. [24] Vaibhav Agarwal, Vaibhav Gupta, Vivasvan Manasvi Vashisht, Kiran Sharma, and Neetu Sharma. Mobile application based cataract detection system. In International Conference on Trends in Electronics and Informatics (ICOEI), pages 780–787, 2019. [25] Yunendah Nur Fuadah, Agung W Setiawan, Tati LR Mengko, et al. Mobile cataract detection using optimal combination of statistical texture analysis. In International Conference on Instrumentation, Communications, Information Technology, and Biomedical Engineering (ICICI-BME), pages 232–236, 2015. [26] Juyel Rana and Syed Md Galib. Cataract detection using smartphone. In International Conference on Electrical Information and Communication Technology (EICT), pages 1–4, 2017. [27] Behnam Askarian, Peter Ho, and Jo Woon Chong. Detecting cataract using smartphones. IEEE Journal of Translational Engineering in Health and Medicine, 9:1–10, 2021. [28] Nattapon Boonarpha, PC Knox, M Batterbury, Claire Rogers, and Y Zheng. Comparison of optic disc images from two smartphone-based imaging systems in glaucoma patients and suspects. Ophthalmology Journal, 2018. [29] Rachel LZ Goh, Yu Xiang George Kong, Colm McAlinden, John Liu, Jonathan G Crowston, and Simon E Skalicky. Objective assessment of activity limitation in glaucoma with smartphone virtual reality goggles: a pilot study. Translational Vision Science & Technology, 7(1):10–10, 2018. [30] Nilesh Kumar, Bandello Francesco, and Ashish Sharma. Smartphone-based gonio-imaging: A novel addition to glaucoma screening tools. Journal of Glaucoma, 28(9):e149–e150, 2019. [31] Fei Li, Diping Song, Han Chen, Jian Xiong, Xingyi Li, Hua Zhong, Guangxian Tang, Sujie Fan, Dennis SC Lam, Weihua Pan, et al. Development and clinical deployment of a smartphone-based visual field deep learning system for glaucoma detection. npj Digital Medicine, 3(1):1–8, 2020. [32] José Martins, Jaime S Cardoso, and Filipe Soares. Offline computer-aided diagnosis for glaucoma detection using fundus images targeted at mobile devices. Computer Methods and Programs in Biomedicine, 192:105341, 2020. [33] Lauren C LaMonica, Mahesh K Bhardwaj, Nicola L Hawley, Take Naseri, Muagatutia Sefuiva Reupena, Michael L Cooper, Paul R Cotran, Shiyoung Roh, and David J Ramsey. Remote screening for optic nerve cupping using smartphone-based non-mydriatic fundus photography. Journal of Glaucoma, 30(1):58, 2021. [34] Yasmine Mrad, Yaroub Elloumi, Mohamed Akil, and Mohamed Hédi Bedoui. A fast and accurate method for glaucoma screening from smartphone-captured fundus images. Irbm, 43(4):279–289, 2022. [35] Min Jiang and Guodong Guo. Body weight analysis from human body images. IEEE Transactions on Information Forensics and Security, 14(10):2676–2688, 2019. [36] Min Jiang, Yuanyuan Shang, and Guodong Guo. Computational approach to body mass index estimation from dressed people in 3d space. IET Image Processing, 14(7):1248–1256, 2020. [37] Lingyun Wen and Guodong Guo. A computational approach to body mass index prediction from face images. Image and Vision Computing, 31(5):392–400, 2013. [38] Olivia Affuso, Ligaj Pradhan, Chengcui Zhang, Song Gao, Howard W Wiener, Barbara Gower, Steven B Heymsfield, and David B Allison. A method for measuring human body composition using digital images. PloS one, 13(11):e0206430, 2018. [39] Thippa Reddy Gadekallu, Celestine Iwendi, Chuliang Wei, and Qin Xin. Identification of malnutrition and prediction of bmi from facial images using real-time image processing and machine learning. IET Image Process, 16:647–658, 2021. 18

[40] Dang Khanh Ngan Ho, Wan-Chun Chiu, Jing-Wen Kao, Hsiang-Tung Tseng, Chih-Yuan Yao, Hsiu-Yueh Su, Pin-Hui Wei, Nguyen Quoc Khanh Le, Hung Trong Nguyen, and Jung-Su Chang. Mitigating errors in mobile-based dietary assessments: Effects of a data modification process on the validity of an image-assisted food and nutrition app. Nutrition, 116:112212, 2023. [41] M. K. Hasan, N. Sakib, R. R. Love, and S. I. Ahamed. Analyzing the existing noninvasive hemoglobin measurement techniques. In 2017 IEEE 8th Annual Ubiquitous Computing, Electronics and Mobile Communication Conference (UEMCON), pages 442–448, 2017. [42] Md Kamrul Hasan, Md Hasanul Aziz, Md Ishrak Islam Zarif, Mahmudul Hasan, MMA Hashem, Shion Guha, Richard R Love, and Sheikh Ahamed. Noninvasive hemoglobin level prediction in a mobile phone environment: State of the art review and recommendations. JMIR mHealth and uHealth, 9(4):e16806, 2021. [43] Edward Jay Wang, William Li, Doug Hawkins, Terry Gernsheimer, Colette Norby-Slycord, and Shwetak N Patel. Hemaapp: noninvasive blood screening of hemoglobin using smartphone cameras. In Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing, pages 593–604, 2016. [44] Edward J Wang, William Li, Junyi Zhu, Rajneil Rana, and Shwetak N Patel. Noninvasive hemoglobin measurement using unmodified smartphone camera and white flash. In 2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pages 2333–2336, 2017. [45] Robert G Mannino, David R Myers, Erika A Tyburski, Christina Caruso, Jeanne Boudreaux, Traci Leong, GD Clifford, and Wilbur A Lam. Smartphone app for non-invasive detection of anemia using only patient-sourced photos. Nature communications, 9(1):1–10, 2018. [46] Md Kamrul Hasan, Md Munirul Haque, Riddhiman Adib, Jannatul F Tumpa, Azima Begum, Richard R Love, Young L Kim, and I Ahamed Sheikh. Smarthelp: Smartphone-based hemoglobin level prediction using an artificial neural network. In AMIA Annual Symposium Proceedings, volume 2018, page 535, 2018. [47] Md Kamrul Hasan, Munirul Haque, Nazmus Sakib, Richard Love, and Sheikh I Ahamed. Smartphone-based human hemoglobin level measurement analyzing pixel intensity of a fingertip video on different color spaces. Smart Health, 5:26–39, 2018. [48] Sunanda Das, Abhishek Kesarwani, Mamata Dalui, Dakshina Ranjan Kisku, Bibhash Sen, Suchismita Roy, and Anupam Basu. Smartphone-based non-invasive haemoglobin level estimation by analyzing nail pallor. Biomedical Signal Processing and Control, 85:104959, 2023. [49] Yuwen Chen, Xiaoyan Hu, Yiziting Zhu, Xiang Liu, and Bin Yi. Real-time non-invasive hemoglobin prediction using deep learning-enabled smartphone imaging. BMC Medical Informatics and Decision Making, 24(1):187, 2024. [50] S. Sukaphat, S. Nanthachaiporn, K. Upphaccha, and P. Tantipatrakul. Heart rate measurement on android platform. In 2016 13th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON), pages 1–5, 2016. [51] Rifat Zaman, Chae Ho Cho, Konrad Hartmann-Vaccarezza, Tra Nguyen Phan, Gwonchan Yoon, and Jo Woon Chong. Novel fingertip image-based heart rate detection methods for a smartphone. Sensors, 17(2):358, 2017. [52] Wei Leong Khong, Muralindran Mariappan, and Chee Siang Chong. Contact and non-contact heart beat rate measurement techniques: Challenges and issues. Pertanika Journal of Science & Technology, 29(3), 2021. [53] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762–10774, 2010. 19

[54] T. Pursche, J. Krajewski, and R. Moeller. Video-based heart rate measurement from human faces. In IEEE International Conference on Consumer Electronics (ICCE), pages 544–545, 2012. [55] Sungjun Kwon, Hyunseok Kim, and Kwang Suk Park. Validation of heart rate extraction using video imaging on a built-in camera system of a smartphone. In International conference of the IEEE engineering in medicine and biology society, pages 2174–2177, 2012. [56] H. Monkaresi, R. A. Calvo, and H. Yan. A machine learning approach to improve contactless heart rate monitoring using a webcam. IEEE Journal of Biomedical and Health Informatics, 18(4):1153–1160, 2014. [57] Yong-Poh Yu, P Raveendran, and Chern-Loon Lim. Dynamic heart rate measurements from video sequences. Biomedical Optics Express, 6(7):2466–2480, 2015. [58] Ahmed Osman, Jay Turcot, and Rana El Kaliouby. Supervised learning approach to remote heart rate estimation from facial videos. In IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), volume 1, pages 1–6, 2015. [59] Guha Balakrishnan, Fredo Durand, and John Guttag. Detecting pulse from head motions in video. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3430–3437, 2013. [60] L. Shan and M. Yu. Video-based heart rate measurement using head motion tracking and ICA. In International Congress on Image and Signal Processing (CISP), volume 01, pages 160–164, 2013. [61] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Optics express, 16(26):21434–21445, 2008. [62] Shuchang Xu, Lingyun Sun, and Gustavo Kunde Rohde. Robust efficient estimation of heart rate pulse from video. Biomedical optics express, 5(4):1124–1135, 2014. [63] Lan Wei, Yonghong Tian, Yaowei Wang, Touradj Ebrahimi, and Tiejun Huang. Automatic webcam-based human heart rate measurements using laplacian eigenmap. In Asian Conference on Computer Vision, pages 281–292, 2012. [64] Zhenyue Zhang and Hongyuan Zha. Principal manifolds and nonlinear dimensionality reduction via tangent space alignment. SIAM journal on scientific computing, 26(1):313–338, 2004. [65] Nazia Hameed, Anglia Ruskin, Kamal Abu Hassan, and M Alamgir Hossain. A comprehensive survey on image-based computer aided diagnosis systems for skin cancer. In International Conference on Software, Knowledge, Information Management & Applications (SKIMA), pages 205–214, 2016. [66] Titus Josef Brinker, Achim Hekler, Jochen Sven Utikal, Niels Grabe, Dirk Schadendorf, Joachim Klode, Carola Berking, Theresa Steeb, Alexander H Enk, and Christof Von Kalle. Skin cancer classification using convolutional neural networks: systematic review. Journal of medical Internet research, 20(10):e11936, 2018. [67] Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. nature, 542(7639):115–118, 2017. [68] Titus J Brinker, Achim Hekler, Alexander H Enk, Carola Berking, Sebastian Haferkamp, Axel Hauschild, Michael Weichenthal, Joachim Klode, Dirk Schadendorf, Tim Holland-Letz, et al. Deep neural networks are superior to dermatologists in melanoma image classification. European Journal of Cancer, 119:11–17, 2019. [69] Mrs Jayashree Hajgude, Aishwarya Bhavsar, Harsha Achara, and Nisha Khubchandani. Skin disease detection using image processing with data mining and deep learning. 2019. 20

[70] Hardik Nahata and Satya P Singh. Deep learning solutions for skin cancer detection and diagnosis. In Machine Learning with Health Care Perspective, pages 159–182. 2020. [71] Ni Zhang, Yi-Xin Cai, Yong-Yong Wang, Yi-Tao Tian, Xiao-Li Wang, and Benjamin Badami. Skin cancer diagnosis based on optimized convolutional neural network. Artificial intelligence in medicine, 102:101756, 2020. [72] Md Shahin Ali, Md Sipon Miah, Jahurul Haque, Md Mahbubur Rahman, and Md Khairul Islam. An enhanced technique of skin cancer classification using deep convolutional neural network with transfer learning models. Machine Learning with Applications, 5:100036, 2021. [73] Manoj Kumar, Mohammed Alshehri, Rayed AlGhamdi, Purushottam Sharma, and Vikas Deep. A de-ann inspired skin cancer detection approach using fuzzy c-means clustering. Mobile Networks and Applications, 25:1319–1329, 2020. [74] Tarun Wadhawan, Ning Situ, Keith Lancaster, Xiaojing Yuan, and George Zouridakis. Skinscan©: a portable library for melanoma detection on handheld devices. In IEEE International Symposium on Biomedical Imaging: From Nano to Macro, pages 133–136, 2011. [75] Upender Kalwa, Christopher Legner, Taejoon Kong, and Santosh Pandey. Skin cancer diagnostics with an all-inclusive smartphone application. Symmetry, 11(6):790, 2019. [76] Omar Abuzaghleh, Buket D Barkana, and Miad Faezipour. Automated skin lesion analysis based on color and shape geometry feature set for melanoma early detection and prevention. In IEEE Long Island Systems, Applications and Technology (LISAT) Conference, pages 1–6, 2014. [77] Chaahat Gupta, Naveen Kumar Gondhi, and Parveen Kumar Lehana. Analysis and identification of dermatological diseases using gaussian mixture modeling. IEEE Access, 7:99407– 99427, 2019. [78] Nasim Alamdari, Kouhyar Tavakolian, Minhal Alhashim, and Reza Fazel-Rezai. Detection and classification of acne lesions in acne patients: A mobile application. In IEEE International Conference on Electro Information Technology (EIT), pages 0739–0743, 2016. [79] Rahat Yasir, Md Shariful Islam Nibir, and Nova Ahmed. A skin disease detection system for financially unstable people in developing countries. Global Science and Technology Journal, 3(1):77–93, 2015. [80] Nawal Soliman ALKolifi ALEnezi. A method of skin disease detection using image processing and machine learning. Procedia Computer Science, 163:85–92, 2019. [81] Yasha Hasija, Nikhil Garg, and Soumya Sourav. Automated detection of dermatological disorders through image-processing and machine learning. In International Conference on Intelligent Sustainable Systems (ICISS), pages 1047–1051, 2017. [82] Shuzlina Abdul-Rahman, Ahmad Khairil Norhan, Marina Yusoff, Azlinah Mohamed, and Sofianita Mutalib. Dermatology diagnosis with feature selection methods and artificial neural network. In IEEE-EMBS Conference on Biomedical Engineering and Sciences, pages 371– 376, 2012. [83] M Shamsul Arifin, M Golam Kibria, Adnan Firoze, M Ashraful Amini, and Hong Yan. Dermatological disease diagnosis using color-skin images. In International Conference on Machine Learning and Cybernetics, volume 5, pages 1675–1680, 2012. [84] Md Nafiul Alam, Tamanna Tabassum Khan Munia, Kouhyar Tavakolian, Fartash Vasefi, Nick MacKinnon, and Reza Fazel-Rezai. Automatic detection and severity measurement of eczema using image processing. In Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pages 1365–1368, 2016. [85] Launcelot C De Guzman, Ryan Paolo C Maglaque, Vianca May B Torres, Simon Philippe A Zapido, and Macario O Cordel. Design and evaluation of a multi-model, multi-level artificial neural network for eczema skin lesion detection. In International conference on artificial intelligence, modelling and simulation (AIMS), pages 42–47, 2015. 21

[86] Haijing Wu, Heng Yin, Haipeng Chen, Moyuan Sun, Xiaoqing Liu, Yizhou Yu, Yang Tang, Hai Long, Bo Zhang, Jing Zhang, et al. A deep learning, image based approach for automated diagnosis for inflammatory skin diseases. Annals of translational medicine, 8(9), 2020. [87] Stephanie Chan, Vidhatha Reddy, Bridget Myers, Quinn Thibodeaux, Nicholas Brownstone, and Wilson Liao. Machine learning in dermatology: current applications, opportunities, and limitations. Dermatology and therapy, 10(3):365–386, 2020. [88] Vaibhav Pande, Syed Mohd Ali, Subodh Kumar, and Puneet Goyal. Automated first aid and medication system for burn victims. In 2014 IEEE International Advance Computing Conference (IACC), pages 623–627. IEEE, 2014. [89] Joohi Chauhan, Rahul Goswami, and Puneet Goyal. Using deep learning to classify burnt body parts images for better burns diagnosis. In Sipaim–Miccai Biomedical Workshop, pages 25–32, 2018. [90] Aliyu Abubakar, Hassan Ugail, and Ali Maina Bukar. Assessment of human skin burns: A deep transfer learning approach. Journal of Medical and Biological Engineering, 40(3):321– 333, 2020. [91] Aliyu Abubakar, Mohammed Ajuji, and Ibrahim Usman Yahya. Comparison of deep transfer learning techniques in human skin burns discrimination. Applied System Innovation, 3(2):20, 2020. [92] Hai Son Tran, Thai Hoang Le, and Thuy Thanh Nguyen. The degree of skin burns images recognition using convolutional neural network. Indian J. Sci. Technol, 9(45):1–6, 2016. [93] Aliyu Abubakar, Hassan Ugail, Kirsty M Smith, Ali Maina Bukar, and Ali Elmahmudi. Burns depth assessment using deep learning features. Journal of Medical and Biological Engineering, 40(6):923–933, 2020. [94] Chong Jiao, Kehua Su, Weiguo Xie, and Ziqing Ye. Burn image segmentation based on mask regions with convolutional neural network deep learning framework: more accurate and more convenient. Burns & trauma, 7, 2019. [95] Hao Liu, Keqiang Yue, Siyi Cheng, Wenjun Li, and Zhihui Fu. A framework for automatic burn image segmentation and burn depth diagnosis using deep learning. Computational and Mathematical Methods in Medicine, 2021, 2021. [96] Ashish Bhelonde, Nikhil Didolkar, Shubham Jangale, and Nikhil L. Kulkarni. Flexible wound assessment system for diabetic patient using android smartphone. In 2015 International Conference on Green Computing and Internet of Things (ICGCIoT), pages 466–469, 2015. [97] Lei Wang, Peder C. Pedersen, Diane M. Strong, Bengisu Tulu, Emmanuel Agu, and Ronald Ignotz. Smartphone-based wound assessment system for patients with diabetes. IEEE Transactions on Biomedical Engineering, 62(2):477–488, 2015. [98] Vidyashree Dalya and D.K. Shedge. Design of smartphone-based wound assessment system. In 2016 International Conference on Automatic Control and Dynamic Optimization Techniques (ICACDOT), pages 709–712, 2016. [99] Varun N. Shenoy, Elizabeth Foster, Lauren Aalami, Bakar Majeed, and Oliver Aalami. Deepwound: Automated postoperative wound assessment and surgical site surveillance through convolutional neural networks. In 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 1017–1021, 2018. [100] Kacie Kaile, Kevin Leiva, Jagadeesh Mahadevan, Ramnarayan V., Alonso Miguel Jr., Mohan Vishwanathan, and Anuradha Godavarty. Low-cost smartphone based imaging device to detect subsurface tissue oxygenation of wounds. In Optics and Biophotonics in Low-Resource Settings V, volume 10869, pages 62 – 65, 2019. [101] Chun Hon Lau, Ken Hung-On Yu, Tsz Fung Yip, Luke Yik Fung Luk, Abraham Ka Chung Wai, Tin-Yan Sit, Janet Yuen-Ha Wong, and Joshua Wing Kei Ho. An artificial intelligenceenabled smartphone app for real-time pressure injury assessment. Frontiers in Medical Technology, 4:905074, 2022. 22

[102] A. Mariakakis, Jacob Baudin, Eric Whitmire, Vardhman Mehta, Megan A. Banks, Anthony Law, Lynn Mcgrath, and S. Patel. Pupilscreen: Using smartphones to assess traumatic brain injury. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., 1:81:1–81:27, 2017. [103] Jina Ko, Matthew A Hemphill, David Gabrieli, Leon Wu, Venkata Yelleswarapu, Gladys Lawrence, Wesley Pennycooke, Anup Singh, Dave F Meaney, and David Issadore. Smartphone-enabled optofluidic exosome diagnostic for concussion recovery. Scientific reports, 6(1):31215, 2016. [104] K Anuradha and K Sankaranarayanan. Comparison of feature extraction techniques to classify oral cancers using image processing. International Journal of Application or Innovation in Engineering & Management (IJAIEM), 2(6):456–462, 2013. [105] Vania Firmalino, Afarin Golabgir Anbarani, Delaney Islip, Bofan Song, Ross Uthoff, Thair Takesh, Ron Liang, and Petra E Wilder-Smith. First clinical results: optical smartphonebased oral cancer screening. In LASERS IN SURGERY AND MEDICINE, volume 50, pages S45–S46, 2018. [106] Ross D Uthoff, Bofan Song, Praveen Birur, Moni Abraham Kuriakose, Sumsum Sunny, Amritha Suresh, Sanjana Patrick, Afarin Anbarani, Oliver Spires, Petra Wilder-Smith, et al. Development of a dual-modality, dual-view smartphone-based imaging system for oral cancer detection. In Design and Quality for Biomedical Technologies XI, volume 10486, page 104860V, 2018. [107] Ross D Uthoff, Bofan Song, Sumsum Sunny, Sanjana Patrick, Amritha Suresh, Trupti Kolur, G Keerthi, Oliver Spires, Afarin Anbarani, Petra Wilder-Smith, et al. Point-of-care, smartphone-based, dual-modality, dual-view, oral cancer screening device with neural network classification for low-resource communities. PloS one, 13(12):e0207493, 2018. [108] Huiping Lin, Hanshen Chen, Luxi Weng, Jiaqi Shao, and Jun Lin. Automatic detection of oral cancer in smartphone-based images using deep learning for early diagnosis. Journal of Biomedical Optics, 26(8):086007–086007, 2021. [109] Prachi Shah, Nilanjan Roy, and Pinakin Dhandhukia. Algorithm mediated early detection of oral cancer from image analysis. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 133(1):70–79, 2022. [110] Qiuyun Fu, Yehansen Chen, Zhihang Li, Qianyan Jing, Chuanyu Hu, Han Liu, Jiahao Bao, Yuming Hong, Ting Shi, Kaixiong Li, et al. A deep learning algorithm for detection of oral cavity squamous cell carcinoma from photographic images: A retrospective study. EClinicalMedicine, 27, 2020. [111] Chang Jin Jung, Young Ju Jeon, Jong Yeol Kim, and Keun Ho Kim. Review on the current trends in tongue diagnosis systems. Integrative Medicine Research, 1(1):13–20, 2012. [112] Bob Zhang, Xingzheng Wang, Jane You, and David Zhang. Tongue color analysis for medical application. Evidence-Based Complementary and Alternative Medicine, 2013, 2013. [113] Qi Zhang, Hui-Liang Shang, Jia-jun Zhu, Min-Min Jin, Wen-Xin Wang, and Qing-sheng Kong. A new tongue diagnosis application on android platform. In 2013 IEEE International Conference on Bioinformatics and Biomedicine, pages 334–327, 2013. [114] Ini Ryu and Itiro Siio. Tonguedx: a tongue diagnosis for health care on smartphones. In Proceedings of the 5th Augmented Human International Conference, pages 1–2, 2014. [115] Min-Chun Hu, Guang-Yu Zheng, Yian-Ting Chen, and KC Lan. Automatic tongue diagnosis using a smart phone. In 2014 IEEE International Conference on Systems, Man, and Cybernetics, 2014. [116] Min-Chun Hu, Ming-Hsun Cheng, and Kun-Chan Lan. Color correction parameter estimation on the smartphone and its application to automatic tongue diagnosis. Journal of medical systems, 40(1):1–8, 2016. 23

[117] Min-Chun Hu, Kun-Chan Lan, Wen-Chieh Fang, Yu-Chia Huang, Tsung-Jung Ho, ChunPang Lin, Ming-Hsien Yeh, Paweeya Raknim, Ying-Hsiu Lin, Ming-Hsun Cheng, Yi-Ting He, and Kuo-Chih Tseng. Automated tongue diagnosis on the smartphone and its applications. Computer Methods and Programs in Biomedicine, 174:51–64, 2019. [118] Ratchadaporn Kanawong, Tayo Obafemi-Ajayi, Dahai Liu, Meng Zhang, Dong Xu, and Ye Duan. Tongue image analysis and its mobile app development for health diagnosis. Translational Informatics in Smart Healthcare, pages 99–121, 2017. [119] Zih-Hao Huang, Wei-Cheng Huang, Hsien-Chang Wu, and Wen-Chieh Fang. Tonguemobile: automated tongue segmentation and diagnosis on smartphones. Neural Computing and Applications, 35(28):21259–21274, 2023. [120] Young Ki Choi and K.-O Lee. A research on the developing smartphone application for automatic dental caries detection. Journal of Engineering and Applied Sciences, 13:10468– 10473, 01 2018. [121] Yipeng Zhang, Haofu Liao, Jin Xiao, Nisreen Al Jallad, Oriana Ly-Mapes, and Jiebo Luo. A smartphone-based system for real-time early childhood caries diagnosis. In Medical Ultrasound, and Preterm, Perinatal and Paediatric Image Analysis, pages 233–242. 2020. [122] Behnam Askarian, Fatemehsadat Tabei, Grace Anne Tipton, and Jo Woon Chong. Smartphone-based method for detecting periodontal disease. In 2019 IEEE Healthcare Innovations and Point of Care Technologies,(HI-POCT), pages 53–55, 2019. [123] Wenzhe You, Aimin Hao, Shuai Li, Yong Wang, and Bin Xia. Deep learning-based dental plaque detection on primary teeth: a comparison with clinical assessments. BMC Oral Health, 20:1–7, 2020. [124] Yuan Liang, Hsuan Wei Fan, Zhujun Fang, Leiying Miao, Wen Li, Xuan Zhang, Weibin Sun, Kun Wang, Lei He, and Xiang’Anthony’ Chen. Oralcam: Enabling self-examination and awareness of oral health using a smartphone camera. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, pages 1–13, 2020. [125] Yukai Huang and Chia-Ping Huang. A depth-first search algorithm based otoscope application for real-time otitis media image interpretation. In 18th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2017, Taipei, Taiwan, December 18-20, 2017, pages 170–175, 2017. [126] Hermanus C. Myburgh, Stacy Jose, De Wet Swanepoel, and Claude Laurent. Towards low cost automated smartphone- and cloud-based otitis media diagnosis. Biomedical Signal Processing and Control, 39:34–52, 2018. [127] Yen-Chi Chen, Yuan-Chia Chu, Chii-Yuan Huang, Yen-Ting Lee, Wen-Ya Lee, Chien-Yeh Hsu, Albert C Yang, Wen-Huei Liao, and Yen-Fu Cheng. Smartphone-based artificial intelligence using a transfer learning algorithm for the detection and diagnosis of middle ear diseases: A retrospective deep learning study. EClinicalMedicine, 51, 2022. [128] Behnam Askarian, Seung-Chul Yoo, and Jo Woon Chong. Novel image processing method for detecting strep throat (streptococcal pharyngitis) using smartphone. Sensors, 19(15):3307, 2019. [129] Tae Keun Yoo, Joon Yul Choi, Younil Jang, and Ik Hee Ryu. Toward automated severe pharyngitis detection with smartphone camera using deep learning networks. Computers in biology and medicine, 125:103980, 2020. [130] Aetsoft. Tonsillitis detection app. https://aetsoft.net/products/ tonsillitis-detection-app/, 2014. Accessed: March 24, 2022. [131] Myung Jin Ban, Yunyoung Nam, and Jae Hong Park. Detection of peritonsillar abscess using smartphone-based thermal imaging. Pakistan Journal of Medical Sciences, 33(2):502, 2017. 24

[132] Hong Luo, Deye Yang, Andrew Barszczyk, Naresh Vempala, Jing Wei, Si Jia Wu, Paul Pu Zheng, Genyue Fu, Kang Lee, and Zhong-Ping Feng. Smartphone-based blood pressure measurement using transdermal optical imaging technology. Circulation: Cardiovascular Imaging, 12(8):e008857, 2019. [133] Zhi Zhou, Eliza Yingzi Du, N Luke Thomas, and Edward J Delp. A new human identification method: Sclera recognition. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 42(3):571–583, 2011. [134] Alex Mariakakis, Megan A Banks, Lauren Phillipi, Lei Yu, James Taylor, and Shwetak N Patel. Biliscreen: smartphone-based scleral jaundice monitoring for liver and pancreatic disorders. ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 1(2):1–26, 2017. [135] Mohammad Rafayet Ali, Javier Hernandez, E Ray Dorsey, Ehsan Hoque, and Daniel McDuff. Spatio-temporal attention and magnification for classification of parkinson’s disease from videos collected via the internet. In International Conference on Automatic Face and Gesture Recognition, pages 53–60, 2020. [136] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv, 2014. [137] Farhana Shahid, Wasifur Rahman, M Saifur Rahman, Sharmin Akther Purabi, Ayesha Seddiqa, Moin Mostakim, Farhan Feroz, Tanjir Rashid Soron, Fahmida Hossain, Nabila Khan, et al. Leveraging free-hand sketches for potential screening of ptsd. ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 4(3):1–22, 2020. [138] Andrea Zunino, Pietro Morerio, Andrea Cavallo, Caterina Ansuini, Jessica Podda, Francesca Battaglia, Edvige Veneselli, Cristina Becchio, and Vittorio Murino. Video gesture analysis for autism spectrum disorder detection. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 3421–3426, 2018. [139] Yuan Tian, Xiongkuo Min, Guangtao Zhai, and Zhiyong Gao. Video-based early asd detection via temporal pyramid networks. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 272–277, 2019. [140] Alaric Aroojis, Bhushan Sagade, and Suresh Chand. Usability and reliability of the edinburgh visual gait score in children with spastic cerebral palsy using smartphone slow-motion video technology and a motion analysis application: A pilot study. Indian Journal of Orthopaedics, pages 1–8, 2021. [141] Priyeta Saha, Syed Muhammad Ibne Zulfiker, Tanzima Hashem, and Khandker Aftarul Islam. Dehydration scan: An artificial intelligence assisted smartphone-based system for early detection of dehydration. In International Conference on Pervasive Computing Technologies for Healthcare, pages 289–312. Springer, 2022. [142] Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, pages 4765–4774, 2017. [143] Marco Túlio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should I trust you?": Explaining the predictions of any classifier. In SIGKDD, pages 1135–1144. ACM, 2016.

A

Appendix

25

Table 6: Summary of Various Healthcare Categories Category Eye Diseases

Disease Diabetic Retinopathy (DR)

Retinopathy of Prematurity (ROP)

Cataract

Glaucoma

Malnutrition

Obesity

Hemoglobin Deficiency

Heart Diseases

Heart Rate Variability

Description Diabetic Retinopathy is an eye disease that damages the retina due to diabetes mellitus, often leading to blindness in working adults. Symptoms: Gradually worsening vision, sudden vision loss, floaters, blurred or patchy vision, eye pain or redness, difficulty seeing in the dark. Diagnosis: Diagnosed through stereoscopic fundus photographs. Retinopathy of Prematurity is an eye disease observed in premature infants and is a leading cause of blindness in children, preventable if treated early. Symptoms: Unusual eye movements, white pupils, vision loss. Diagnosis: Diagnosed using fundus images. Cataract is a common eye condition in the elderly that leads to poor vision and eventually blindness. Symptoms: Clouded or blurry vision, difficulty seeing at night, light sensitivity, seeing halos around lights, frequent changes in eyeglass prescription. Diagnosis: Diagnosed through a visual acuity test. Glaucoma is a cause of visual impairment and blindness that is often asymptomatic in its early stages, leading to undiagnosed cases until advanced. Symptoms: Intense eye pain, nausea, vomiting, red eyes, headaches, seeing rings around lights, blurred vision. Diagnosis: Diagnosed using the Goldmann tonometer. Obesity, characterized by excess body weight, is a major health condition leading to various associated health problems. Symptoms: Breathlessness, increased sweating, snoring, difficulty in physical activity, frequent tiredness, joint and back pain. Diagnosis: Diagnosed through anthropometric measures such as weight, height, BMI, and waist-hip ratio. Hemoglobin carries oxygen to the body’s tissues, and a deficiency may signal iron-deficient anemia caused by malnutrition, among other diseases. Symptoms: Pale skin, brittle nails, pica syndrome (eating non-food items), lightheadedness when standing, shortness of breath, sore or inflamed tongue, mouth ulcers. Diagnosis: Diagnosed through blood tests. Heart rate variability refers to the fluctuations in time intervals between adjacent heartbeats. An optimal heart rate variability is indicative of good health, adaptability, and resilience. Symptoms: Fatigue, inflammation. Diagnosis: Diagnosed via an electrocardiogram (ECG). Continued on next page

26

Table 6 – Summary of Various Healthcare Categories [Continued] Disease Description Skin Cancer Skin cancer is a type of cancer that develops in the skin, the body’s largest organ. It includes basal cell carcinoma, squamous cell carcinoma, and melanoma, with melanoma being the most unpredictable and dangerous. Symptoms: A new spot on the skin, changes in the size, shape, or color of an existing spot, appearance of spots that are itchy or painful, non-healing sores that bleed or develop a crust. Diagnosis: Diagnosed using a skin biopsy with histopathology. Inflammatory Inflammatory skin diseases, such as eczema, hives, psoriSkin Diseases asis, atopic dermatitis, herpes, are prevalent in the modern population. These may be triggered by allergens, extreme temperatures and even stress. Despite their prevalence, these diseases are susceptible to misdiagnosis by inexperienced dermatologists. Symptoms: Smooth or scaly skin, skin may itch, burn or sting, flat or raised skin, skin redness, warmth in the affected skin area, blisters or pimples. Diagnosis: Diagnosed using histopathology. Skin Burns Skin burns are injuries caused by heat, electricity, radiation, Injury chemicals, etc. Burns are classified into first, second, or third Detection degree based on their severity. Symptoms: Blisters, pain, swelling, red, white, or charred skin. Diagnosis: Diagnosed using a punch biopsy of burn tissue with subsequent histological analysis. Wounds A wound can be considered an injury to living tissue caused by a cut, bump, or other impacts. It can be a symptom of other critical diseases such as diabetes or ulcers. Symptoms: Fever, swelling, redness, bleeding. Diagnosis: Can be detected via naked eyes. Brain Injuries Brain injuries are damage to the brain caused by external forces, such as a blow or impact, with concussions being a common result from rapid brain movement within the skull. Symptoms: Amnesia, headache, loss of consciousness, balance problems, nausea. Diagnosis: Diagnosed using MRI or CT scans. Oral Cancer Oral cancer affects the mouth, including the lips, tongue, and other oral tissues, and is associated with high mortality Oral Diseases rates, particularly if diagnosis is delayed. Symptoms: Lump in neck, loose teeth, sore lips, difficult swallowing, unexplained weight loss. Diagnosis: Diagnosed using tissue biopsy with histological assessment. Tongue Diseases Tongue diseases refer to conditions affecting the tongue, where changes in its coating and color can indicate underlying health issues, such as a purple tongue signaling high cholesterol or chronic bronchitis. Symptoms: Pain, bumps on the surface of the tongue and changes to its color or texture. Diagnosis: Diagnosed via histological examination. Continued on next page Category Skin Diseases

27

Category

ENT

Others

Table 6 – Summary of Various Healthcare Categories [Continued] Disease Description Dental & PeriDental and periodontal diseases refer to conditions affectodontal Diseases ing the teeth and the supporting structures, primarily caused by infections and inflammation of the gums and bone that surround and support the teeth Symptoms: Severe pain, infection, tooth loss, inflammation in surrounding tissues. Diagnosis: Diagnosed via periodontal probing, measurement of clinical attachment loss, bone loss, cone beam computed tomography, quantitative polymerase chain reaction tests, etc. Ear Infection An ear infection is an inflammation of the ear, often caused by bacteria or viruses. Otitis media, a common type, can lead to hearing loss and impact a child’s learning ability Symptoms: Irritability, difficulty in sleeping or staying asleep, fever, fluid draining from ear(s), loss of balance, hearing difficulties, ear pain. Diagnosis: Diagnosed using an otoscope. Throat Diseases Throat diseases involve infections or inflammation of the throat, such as streptococcal pharyngitis (strep throat), a bacterial infection affecting the throat and tonsils. Symptoms: Fever, pain when swallowing, sore throat, red and swollen tonsils, white patches or streaks of pus on the tonsils. Diagnosis: Diagnosed using a throat culture test. Blood Pressure High blood pressure (hypertension) is a condition where the (BP) force of blood against the artery walls is consistently too high, leading to complications such as heart failure, vision loss, stroke, and kidney diseases. Symptoms: Headaches, blurred vision, chest pain. Diagnosis: Measured using a cuff-based method (sphygmomanometer). Jaundice Jaundice is a condition caused by an accumulation of bilirubin in the blood, often due to liver diseases, leading to yellowing of the skin, eyes, and mucous membranes. Symptoms: Dark urine, abdominal pain, fever, nausea, yellow skin, pale stools, chills. Diagnosis: Diagnosed using a bilirubin blood test. Post-traumatic PTSD is a mental health disorder that occurs after experiStress Disorder encing or witnessing traumatic events, leading to flashbacks, (PTSD) nightmares, and severe anxiety. Symptoms: Flashbacks, nightmares and severe anxiety. Diagnosis: Diagnosed using the CAPS-5, a 30-item structured interview. Mental Health Mental health conditions involve psychological disorders, Diseases including depression, anxiety, and other disorders affecting emotional well-being and mental functioning. Symptoms: poor concentration, guilt or low self-worth, hopelessness, suicidal thoughts etc. Diagnosis: Diagnosed through therapy sessions and psychological assessments. Gait Abnormali- Gait abnormalities refer to unusual walking patterns, often ties seen in conditions like cerebral palsy. Symptoms: Walking with head and neck bent, dragging feet, shuffling, or irregular jerky movements. Diagnosis: Diagnosed using the Edinburgh Visual Gait Score (EVGS). Continued on next page

28

Category

Table 6 – Summary of Various Healthcare Categories [Continued] Disease Description Dehydration Dehydration occurs when the body loses more fluids than it consumes, leading to an imbalance in electrolytes and affecting normal bodily functions. Symptoms: Headache, confusion, dizziness, fatigue, sunken eyes, dry mouth, dry cough. Diagnosis: Diagnosed using plasma osmolality measurements.

29