Multi-regional Multiparametric Deep Learning Radiomics for Diagnosis of Clinically Significant Prostate Cancer.

Multi-regional Multiparametric Deep Learning Radiomics for Diagnosis of Clinically Significant Prostate Cancer - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Imaging Inform Med . 2025 Aug 29;39(3):2123–2135. doi: 10.1007/s10278-025-01551-1 Multi-regional Multiparametric Deep Learning Radiomics for Diagnosis of Clinically Significant Prostate Cancer Xijun Liu 1, 2 , Rongzong Liu 3 , Haihao He 1, 2 , Yifei Yan 1, 2 , Limin Zhang 3, ✉ , Qi Zhang 1, 2, ✉ Author information Article notes Copyright and License information 1 The SMART (Smart Medicine and AI-Based Radiology Technology) Lab, School of Communication and Information Engineering, Shanghai University, Shanghai, China 2 Shanghai Institute for Advanced Communication and Data Science, Shanghai University, Shanghai, China 3 Department of Urology, Huashan Hospital, Fudan University, Shanghai, China ✉ Corresponding author. Received 2025 Feb 24; Revised 2025 May 12; Accepted 2025 May 13; Collection date 2026 Jun. © The Author(s) under exclusive licence to Society for Imaging Informatics in Medicine 2025. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC13230347  PMID: 40879858 Abstract Non-invasive and precise identification of clinically significant prostate cancer (csPCa) is essential for the management of prostatic diseases. Our study introduces a novel and interpretable diagnostic method for csPCa, leveraging multi-regional, multiparametric deep learning radiomics based on magnetic resonance imaging (MRI). The prostate regions, including the peripheral zone (PZ) and transition zone (TZ), are automatically segmented using a deep learning framework that combines convolutional neural networks and transformers to generate region-specific masks. Radiomics features are then extracted and selected from multiparametric MRI at the PZ, TZ, and their combined area to develop a multi-regional multiparametric radiomics diagnostic model. Feature contributions are quantified to enhance the model’s interpretability and assess the importance of different imaging parameters across various regions. The multi-regional model substantially outperforms single-region models, achieving an optimal area under the curve (AUC) of 0.903 on the internal test set, and an AUC of 0.881 on the external test set. Comparison with other methods demonstrates that our proposed approach exhibits superior performance. Features from diffusion-weighted imaging and apparent diffusion coefficient play a crucial role in csPCa diagnosis, with contribution degrees of 53.28% and 39.52%, respectively. We introduce an interpretable, multi-regional, multiparametric diagnostic model for csPCa using deep learning radiomics. By integrating features from various zones, our model improves diagnostic accuracy and provides clear insights into the key imaging parameters, offering strong potential for clinical applications in csPCa management. Keywords: Clinically significant prostate cancer, Multiparametric magnetic resonance imaging, Radiomics, Multi-regional Introduction Prostate cancer (PCa) accounts for 14.2% of all newly diagnosed cancers in men, and it has the second-highest incidence rate after lung cancer [ 1 , 2 ]. Prostate cancer (PCa) is commonly classified into clinically significant (csPCa) and clinically insignificant PCa (cisPCa), primarily based on the Gleason score determined through pathological assessment [ 3 ]. Tumors with a Gleason score ≥ 7 are defined as csPCa, whereas those with lower scores are considered cisPCa. csPCa is highly invasive and carries a significant risk of mortality, requiring timely diagnosis and active treatment. In contrast, cisPCa and benign prostate diseases, collectively referred to as non-csPCa, are typically managed through active surveillance and conservative treatment [ 4 ]. Prostate biopsy is commonly used as a pathological method to discriminate between csPCa and non-csPCa [ 5 ]. Nevertheless, as an invasive procedure, biopsy carries risks such as infection and bleeding and often causes significant discomfort for patients, which limits its broader use in early screening [ 6 ]. However, traditional non-invasive diagnostic methods, including digital rectal examination and prostate-specific antigen testing, have limited sensitivity and specificity, often resulting in high false-positive and false-negative rates [ 7 , 8 ]. Therefore, there is an urgent need to develop accurate, non-invasive, and efficient diagnostic methods that can ensure reliable detection of csPCa while reducing the burden on patients [ 9 ]. To address this need, magnetic resonance imaging (MRI) has emerged as a key diagnostic tool for csPCa. The MRI technique with multiple imaging sequences called multiparametric MRI (mpMRI) includes T2-weighted imaging (T2 WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC), providing complementary anatomical and functional information about prostate [ 10 , 11 ]. Specifically, T2 WI offers high-resolution structural images that highlight the morphological characteristics of lesions [ 12 ]. DWI assesses the diffusion properties of water molecules to detect microenvironmental changes within tissues, while ADC quantitatively complements DWI by effectively differentiating between benign and malignant tissues [ 13 ]. Despite its clear advantages, the diagnostic accuracy of mpMRI depends heavily on the expertise of radiologists, leading to variability and inconsistency in diagnostic results. Hence, advancing technologies to support automated mpMRI interpretation has become a key focus of research aimed at improving diagnostic accuracy and reliability [ 14 , 15 ]. The anatomical complexity of the prostate poses another significant challenge for mpMRI-based csPCa diagnosis. The prostate can be histologically divided into three main regions, including the peripheral zone (PZ), the transition zone (TZ), and the central zone, with the PZ and TZ being the primary sites for csPCa development [ 16 ]. Studies indicate that 70–80% of PCa cases originate in the PZ, while 20–25% occur in the TZ [ 17 ]. The MRI signals of the PCa at PZ and TZ differ noticeably, making accurate diagnosis both challenging and crucial for appropriate management [ 18 , 19 ]. In the PZ, PCa often appears as a region of low signal intensity on T2 WI due to its relatively lower water content and higher cell density compared to the surrounding normal tissue. This hypointensity can be a key indicator of malignancy in this zone. On DWI, PZ tumors tend to exhibit high signal intensity because of the restricted diffusion caused by increased cellular density. Benign prostatic hyperplasia (BPH), which is common in the TZ, can cause low signal in T2 WI and high signal in DWI that mimic those of PCa, which complicates the detection of PCa in the TZ [ 13 ]. Consequently, an effective diagnostic system should not only integrate anatomical and functional information delivered by multiparametric imaging but also account for the unique imaging characteristics of each zone to achieve precise, reliable, and comprehensive diagnosis. Radiomics has emerged as a promising tool for addressing these challenges, offering a quantitative analysis of imaging features linked to pathological information to support clinical decision-making [ 20 ]. However, radiomics methods often require radiologists to manually delineate regions of interest (ROIs), which is both time-consuming and prone to human error. Hence, the deep learning–based automatic segmentation of ROIs has emerged to improve accuracy, reproducibility, and efficiency. In the radiomics process, using deep learning for segmentation, termed into the category of the deep learning radiomics (DLR), expands the possibilities for diagnosing csPCa [ 16 ]. For example, Zheng et al. proposed an automatic deep-radiomics framework that uses deep learning to automatically segment the prostate and suspicious lesion regions, combining imaging features and clinical data for the diagnosis and stratification of PCa [ 21 ]. Another type of the DLR methods involves directly predicting csPCa with deep learning in an end-to-end manner [ 22 ]. For instance, Saha et al. proposed a DLR framework that integrates anatomical priors with attention mechanisms, substantially improving diagnostic sensitivity while reducing false-positive rates [ 23 ]. Despite these advancements, current DLR research still faces several limitations. First, existing DLR studies mostly focus on single-region analysis of the prostate whole gland (WG), without fully exploring the unique and complementary characteristics of the PZ and TZ for csPCa diagnosis [ 24 ]. Second, these models often struggle to handle high-dimensional features and redundancy when integrating multiparametric images [ 20 , 23 ]. The third critical limitation is the “black-box” nature of many deep learning models, which hinders their adoption in clinical practice, especially in multi-regional and multiparametric settings. To address these issues, we propose a multi-regional multiparametric DLR framework that combines advanced segmentation, feature extraction, and model interpretability methods. Specifically, the framework employs an automatic zonal segmentation method by combining a convolutional neural network (CNN) and transformer for automatic segmentation of both PZ and TZ in MRI images, followed by the extraction of multiparametric radiomics features, including T2 WI, DWI, and ADC, from each region of PZ, TZ, or WG [ 25 ]. Independent diagnostic models are constructed for the PZ, TZ, and WG regions, and a model integration strategy is introduced to enhance overall diagnostic performance. Additionally, the contributions of features from various imaging sequences and prostate regions to diagnostic outcomes are quantified to result in a transparent and interpretable diagnostic framework. Our study introduces three major innovations compared to existing research: We employ a deep learning architecture that integrates CNN and transformer to achieve precise automatic segmentation of the prostatic PZ and TZ for zonal feature extraction. We propose a multi-regional multiparametric diagnostic model by independently constructing diagnostic models for PZ, TZ, and WG, followed by designing a multi-regional feature integration strategy to fully exploit the specificity and complementary information across regions. We introduce an interpretable feature contribution analysis framework to systematically evaluate the diagnostic contributions of multiparametric imaging features, including T2 WI, DWI, and ADC, across multiple regions of PZ, TZ, and WG. Methods Overall Architecture of Radiomics Method We propose a multi-regional multiparametric diagnostic framework for diagnosing csPCa, integrating a deep learning segmentation model into radiomics analysis to deliver an accurate method for prostate MRI image analysis. As illustrated in Fig. 1 , the process begins with the zonal segmentation of the PZ and TZ in T2 WI by adopting the convolution coupled transformer U-shaped network (CCT-Unet), which combines CNN and transformer to leverage both local and global information and generate high-quality regional masks [ 25 ]. These masks are subsequently mapped onto aligned DWI and ADC images, ensuring consistency across multiparametric imaging data and providing precise regional definitions for radiomics feature extraction. Radiomics features, including morphological, statistical, and textural characteristics, are then computed from the PZ, TZ, and WG. These features are initially employed to construct single-regional diagnostic models, which are further integrated using integrated learning methods to form a multi-regional diagnostic model. Finally, the clinical utility and interpretability of the proposed framework are assessed through the receiver operating characteristic (ROC) curves and interpretative feature importance analysis. Fig. 1. Open in a new tab Workflow of our multi-regional multiparametric deep learning radiomics framework for csPCa diagnosis Preprocessing of Multiparametric Images For the diagnosis of csPCa, mpMRI images comprising T2 WI, DWI, and ADC sequences play a critical role due to their ability to provide complementary structural and functional information. However, differences in acquisition parameters among these sequences result in inconsistencies in spatial resolutions, fields of view, and acquisition matrices. For example, the spatial resolution of ADC images is often lower than that of T2 WI images, which can hinder direct comparisons and analyses across sequences. Therefore, preprocessing is essential to ensure consistency and alignment among multiparametric imaging data [ 26 ]. The primary goal of preprocessing is to resample DWI and ADC images to match the spatial resolution and spatial coordinate system of T2 WI images. First, the nearest-neighbor interpolation is applied to adjust the spatial resolution of DWI and ADC images to align with T2 WI images. Next, the fields of view and orientations of the input images are adjusted based on the reference parameters of T2 WI, ensuring proper alignment in a unified spatial coordinate system. Finally, rigid transformations, including rotation and translation, are applied to minimize spatial discrepancies among the sequences. By standardizing the spatial resolution and aligning the coordinate systems of multiparametric images, these preprocessing steps provide a unified spatial framework for subsequent feature extraction and analysis. Typical examples of mpMRI of csPCa and non-csPCa after standardization and alignment are shown in Fig. 2 . Fig. 2. Open in a new tab Prostate mpMRI of (a) csPCa and (b) non-csPCa. From left to right, the three columns show T2 WI, DWI, and ADC, respectively Prostate Zonal Segmentation The segmentation of PZ and TZ is based on our previous work called the CCT-Unet model, which synergistically combines the local feature extraction capabilities of CNN with the global context modeling strength of transformer [ 16 ]. Here, we employ the CCT-Unet for PZ and TZ segmentation to handle the blurred boundaries and large intensity variations in prostate T2 WI images. By incorporating a convolutional embedding module, the CCT-Unet retains detailed input features during the encoding stage. Simultaneously, its cross-shaped window self-attention mechanism effectively captures long-range global dependencies, enabling more precise segmentation with clearer regional contours. To further address semantic mismatches caused by differences in feature representations between CNN and transformer, the model introduces a feature alignment module. This module enhances feature fusion during the encoding and decoding processes, substantially improving segmentation performance. By jointly modeling local details and global context, the CCT-Unet excels in zonal segmentation of PZ and TZ in T2 WI, providing reliable input data for the subsequent diagnostic framework. Multi-regional Multiparametric Feature Extraction Radiomics features are extracted from T2 WI, DWI, and ADC images of the PZ, TZ, and WG. These features include morphological features, intensity features, and texture features. Morphological features quantify the three-dimensional geometry and boundary characteristics of prostate regions, providing insights into lesion shape and growth patterns. In addition to fundamental measures such as region volume, surface area, maximum two-dimensional slice diameter, longest three-dimensional diameter, and surface-to-volume ratio, more advanced descriptors like compactness (how closely the shape approximates a sphere), elongation (ratio of principal axes), and solidity (volume relative to its convex hull) are computed to capture boundary irregularity and anisotropic growth. These shape metrics have been associated with tumor aggressiveness and aid in distinguishing clinically significant lesions from indolent ones. Intensity features characterize the distribution of voxel values within each region of interest by analyzing grayscale histograms [ 27 ]. Beyond basic measures, such as mean, median, and variance, additional statistics enrich the depiction of tissue heterogeneity and contrast. These include skewness and kurtosis to assess distribution asymmetry and peakedness, entropy to quantify randomness, energy (sum of squared intensities reflecting uniformity), and percentile-based metrics (for example 10 th and 90 th percentiles). Texture features explore spatial relationships of gray levels using statistical matrices. The gray-level co-occurrence matrix (GLCM) assesses pairwise pixel intensity dependencies, while the gray-level run-length matrix (GLRLM) quantifies lengths of homogeneous intensity runs. To capture both fine and coarse textural patterns, we apply wavelet filtering which decomposes images into frequency sub-bands and Laplacian of Gaussian filtering which enhances edge and blob detection at multiple scales. From each filtered image, we extract the complete set of morphological, intensity, and texture descriptors to ensure a comprehensive multi-scale representation of prostate tissue [ 28 , 29 ]. Through the integration of morphological, intensity, and texture features, high-dimensional radiomics features are extracted for the PZ, TZ, and WG from T2 WI, DWI, and ADC images. These features provide a detailed and robust description of the prostate regions, encompassing their geometric structure, grayscale distribution, and textural complexity, thereby forming a strong data foundation for subsequent feature selection and classification modeling. Feature Selection and Classification To address redundancy and overfitting caused by high-dimensional radiomics features, we propose a hybrid algorithm which combines the least absolute shrinkage and selection operator (LASSO) and Shapley additive explanation (SHAP) [ 30 , 31 ]. The LASSO employs L1 regularization to shrink the coefficients of less important features to zero, thereby selecting the most contributive features for the diagnostic task. The optimization objective of LASSO is as follows: β ^ = argmin β 1 2 M ∑ m = 1 M y m - X m β 2 + λ ∑ i = 1 I β i 1 where y m represents the label of the m- th sample, X m is the feature vector of the m- th sample, β i is the coefficient of the i- th feature, λ is the regularization parameter optimized through cross-validation, and M and I are the total numbers of samples and features, respectively. The LASSO effectively reduces feature dimensionality and improves the generalizability of the model. We further use the SHAP method for feature selection by evaluating the contribution of each feature to the diagnostic task. The SHAP value ϕ i of each feature i is calculated as ϕ i = ∑ S ∈ N / i S ! N - S - 1 ! N ! f S ∪ i - f S 2 where N is the set of all features, S is a subset of features excluding feature i , and f ( S ) is the model prediction based on the feature subset S . The formula calculates the contribution of feature i by averaging over all possible subsets S , reflecting the model’s performance improvement when feature i is added. By combining the advantages of LASSO and SHAP, the final feature selection process proceeds as follows: first, the SHAP values of all input features are computed and ranked to generate a feature importance list; the ranked feature set is then input into the LASSO model, where key features are selected iteratively through regularization. The intersection of the LASSO-selected and SHAP-ranked features is taken to ensure the stability and reliability of the selected features. To prevent potentially useful features from being overlooked, those with slightly lower contributions below the intersection threshold are reconsidered for further optimization. Based on the selected feature set, the support vector machine (SVM), the linear discriminant analysis (LDA), and the logistic regression (LR) are employed as classifiers to build the diagnostic models. To further improve diagnostic performance, a region integration strategy is proposed. This strategy integrates the predictions of single-regional (PZ, TZ, and WG) models, combining local and global features to construct a multi-regional radiomics model. The prediction probability of the integrated model is calculated using a weighted averaging approach as follows: P Int = 1 3 P PZ + P TZ + P WG 3 where P Int represents the final prediction probability of the integrated model, and P PZ , P TZ , and P WG represent the prediction probabilities of the single-regional models. Diagnostic Model Evaluation and Feature Contribution Quantification To comprehensively evaluate the performance of the classification models, metrics including area under the receiver operating characteristic curve (AUC), sensitivity (SEN), specificity (SPE), accuracy (ACC), and Youden index (YI) are used [ 32 ]. AUC measures the overall classification performance, SEN and SPE assess the model’s ability to recognize positive and negative cases, ACC reflects overall prediction accuracy, and YI serves as a comprehensive indicator balancing sensitivity and specificity. To enhance the interpretability of the model’s predictions and quantify the importance of individual features, we again employ SHAP method to analyze the feature contributions in the radiomics diagnosis models [ 33 , 34 ]. The contribution of each feature to the model’s prediction is calculated by the SHAP method, thereby revealing its importance within the model [ 35 ]. Specifically, the contribution of each feature i is computed using Eq. ( 2 ). By calculating the SHAP values of all features for a single sample, the local contribution of each feature to that sample’s prediction can be determined. To obtain the global contribution of a feature, the absolute values of its SHAP values across all samples are averaged, as shown in the following formula: ϕ i ¯ = 1 J ∑ j = 1 J ϕ i , j 4 where ϕ i ¯ denotes the global contribution of feature i , ϕ ij is the SHAP value of feature i for sample j in the test set, and J is the total number of test set samples. To evaluate the relative importance of different MRI parameters (T2 WI, DWI, and ADC) in diagnosing csPCa, the cumulative SHAP values of all features associated with a specific imaging parameter are summed and averaged to compute the parameter-level contribution, defined as the total parametric SHAP value T param : T param = ∑ i ∈ P ϕ i I 5 which represents the contribution of a specific MRI parameter such as T2 WI, DWI, or ADC, and P is the feature set of the specific parameter. To quantify the feature contribution, the contribution degree of T2 WI, DWI, and ADC was calculated by dividing their corresponding SHAP values by the sum of the SHAP values for these three parameters. Through the above methods, the contributions of individual features and MRI parameters to the model can be quantified, revealing the relative importance of different imaging sequences in diagnosing csPCa. Results Datasets This study utilized two mpMRI datasets, namely the ProstateX dataset and the Huashan dataset, which served as the internal and external datasets, respectively. Data collection and processing adhered to the ethical principles outlined in the Declaration of Helsinki [ 36 ]. Detailed information about these datasets was summarized in Table 1 . Each patient was diagnosed by biopsy histopathology as the reference diagnosis and classified as csPCa or non-csPCa. Table 1. Summary of the internal and external datasets Datasets Internal (ProstateX) External (Huashan) Equipment type 2 4 Magnetic strength 3 T 3 T Spatial resolution of T2 WI 3.0 × 0.5 × 0.5 mm 3 3.6 × 0.8 × 0.8 mm 3 Number of patients 200 92 Number of csPCa patients 65 24 Number of non-csPCa patients 135 68 Open in a new tab ProstateX The ProstateX dataset was obtained from the Radboud University Medical Center and was publicly available through The Cancer Imaging Archive [ 37 ]. This dataset, employed in the SPIE-AAPM-NCI ProstateX challenge, comprises T2 WI, DWI, and ADC sequences. To protect patient privacy, all data were anonymized prior to release. The region annotations for the PZ and TZ were provided by Liu et al. [ 38 ]. Huashan The Huashan dataset was a private dataset collected from Huashan Hospital, Fudan University, Shanghai, and it includes mpMRI data from 92 patients. All patient data were de-identified to ensure privacy. Each patient underwent standard MRI scans, including T2 WI, DWI, and ADC imaging. The PZ and TZ regions were initially annotated by an attending urologist with 6 years of clinical experience using ITK-Snap software. To enhance accuracy and consistency, these annotations were subsequently reviewed and refined by an associate chief urologist with 12 years of clinical experience. Experimental Setting The experiments were conducted on a server equipped with an Intel(R) Core(TM) i7-10870H 2.20 GHz CPU, alongside the NVIDIA GeForce RTX 3090 24 GB. In our experiments, both segmentation and classification tasks were involved. For the internal dataset, the training and test data for both the segmentation and classification models were consistent, with 160 samples used for training and 40 samples used for internally testing the models’ performance. Given the large domain difference between the internal and external datasets, after training the segmentation model on the internal dataset, it was fine-tuned on the external dataset to improve segmentation performance. Meanwhile, the classification model was re-trained on the external dataset. The external dataset was divided in a ratio of 2:1 for fine-tuning/re-training and externally testing the segmentation and classification models. Results on the Internal Dataset The experimental results, as presented in Table 2 , offered a comparative analysis of the performance metrics for three classifiers, namely LDA, LR, and SVM, across various regions and when employing a region integration strategy. Overall, the region integration model demonstrated substantially better performance than the single-regional models across all classifiers, particularly in key metrics such as AUC, SEN, and SPE. Table 2. Experimental results on the internal dataset. The optimal result is in bold Model Region AUC SEN SPE ACC YI LDA PZ 0.594 ± 0.186 0.759 ± 0.243 0.668 ± 0.261 0.681 ± 0.075 0.427 ± 0.218 TZ 0.858 ± 0.082 0.975 ± 0.070 0.787 ± 0.127 0.815 ± 0.064 0.763 ± 0.123 WG 0.760 ± 0.109 0.937 ± 0.111 0.706 ± 0.158 0.791 ± 0.063 0.643 ± 0.152 Region integration 0.903 ± 0.064 0.990 ± 0.043 0.819 ± 0.091 0.879 ± 0.053 0.809 ± 0.089 LR PZ 0.596 ± 0.189 0.772 ± 0.247 0.661 ± 0.260 0.707 ± 0.074 0.433 ± 0.222 TZ 0.872 ± 0.073 0.982 ± 0.061 0.803 ± 0.116 0.844 ± 0.058 0.785 ± 0.115 WG 0.748 ± 0.107 0.933 ± 0.117 0.687 ± 0.164 0.785 ± 0.066 0.620 ± 0.153 Region integration 0.886 ± 0.062 0.992 ± 0.041 0.802 ± 0.091 0.899 ± 0.043 0.793 ± 0.084 SVM PZ 0.584 ± 0.198 0.751 ± 0.256 0.674 ± 0.264 0.707 ± 0.077 0.425 ± 0.229 TZ 0.862 ± 0.078 0.973 ± 0.076 0.790 ± 0.125 0.819 ± 0.062 0.763 ± 0.122 WG 0.755 ± 0.109 0.933 ± 0.111 0.705 ± 0.165 0.787 ± 0.064 0.638 ± 0.153 Region integration 0.891 ± 0.058 0.980 ± 0.066 0.828 ± 0.097 0.899 ± 0.046 0.807 ± 0.087 Open in a new tab For the single-regional models, the TZ region generally outperformed the PZ and WG regions, as demonstrated across multiple metrics. For instance, with LR, the TZ region achieved an AUC of 0.872, an SEN of 0.982, and an SPE of 0.803, showcasing its strong ability to identify both positive and negative samples. However, while the TZ region exhibited high diagnostic performance, the region integration model further improved upon these results by effectively combining features from multiple regions. For example, in the LDA model, the region integration achieved an AUC of 0.903, surpassing the AUC of the TZ region alone (0.858). Furthermore, the ACC reached 0.879, indicating that the combination of regional features enhanced the overall classification capability by leveraging complementary information from the PZ and TZ regions. The results for LR and SVM classifiers further emphasized the benefits of the region integration method. In the LR model, the AUC of the TZ region was 0.872, while with the region integration, the AUC increased to 0.886, along with a notably high SEN of 0.992. This improvement highlighted the ability of the integration model to enhance the detection of positive samples, which was critical for accurate diagnosis. Similarly, in the SVM model, the PZ region showed weaker performance with an AUC of 0.584. However, after region integration, the AUC improved substantially to 0.891, and the SEN reached 0.980 with an SPE of 0.828. These results illustrated that the region integration approach not only improved sensitivity but also effectively balanced the false-positive rate, leading to more robust and reliable diagnostic outcomes. The ROC curves in Fig. 3 a further validated the advantages of the region integration model. Across all three classifiers, the ROC curves of the region integration moved closer to the top-left corner, indicating a larger AUC value compared to single-regional models. Among them, the LDA achieved the best performance, with an AUC of 0.903 and a curve nearing the top-left corner, highlighting its superior discriminative ability. The AUCs for SVM and LR with region integration were 0.891 and 0.886, respectively, which, while slightly lower than LDA, still substantially outperformed the single-regional counterparts. Fig. 3. Open in a new tab Diagnostic performance of the region integration model on internal ( a ) and external ( b ) datasets, shown by ROC curves for LDA, LR, and SVM classifiers Obviously, although the individual regional models performed well in specific indicators, especially the TZ region performed relatively superior when used alone, the region integration model achieved a significant improvement in overall performance by fusing the complementary features of PZ, TZ, and WG. The region integration not only achieved the highest value in terms of the AUC index but also achieved a better balance between sensitivity and specificity, which reflected the effectiveness and robustness of multi-regional feature fusion in csPCa diagnosis. Results on the External Dataset Table 3 shows the experimental results of the LDA, LR, and SVM classifiers across a single region (PZ, TZ, or WG) or under the region integration scenario. Overall, the region integration model outperformed single-regional models across all classifiers in terms of AUC, SEN, and SPE, demonstrating large performance improvements on the external dataset similarly as on the internal dataset. Table 3. Experimental results on the external dataset. The optimal result is in bold Model Region AUC SEN SPE ACC YI LDA PZ 0.698 ± 0.111 0.810 ± 0.149 0.701 ± 0.166 0.727 ± 0.111 0.512 ± 0.157 TZ 0.725 ± 0.095 0.971 ± 0.083 0.607 ± 0.126 0.690 ± 0.090 0.578 ± 0.107 WG 0.765 ± 0.114 0.719 ± 0.189 0.858 ± 0.176 0.818 ± 0.111 0.577 ± 0.155 Region integration 0.869 ± 0.069 0.913 ± 0.124 0.791 ± 0.147 0.794 ± 0.078 0.704 ± 0.119 LR PZ 0.693 ± 0.108 0.828 ± 0.150 0.669 ± 0.177 0.708 ± 0.118 0.497 ± 0.160 TZ 0.722 ± 0.094 0.978 ± 0.067 0.607 ± 0.120 0.705 ± 0.086 0.585 ± 0.108 WG 0.683 ± 0.109 0.829 ± 0.153 0.658 ± 0.170 0.701 ± 0.119 0.487 ± 0.154 Region integration 0.881 ± 0.055 0.977 ± 0.062 0.781 ± 0.099 0.803 ± 0.070 0.758 ± 0.088 SVM PZ 0.675 ± 0.105 0.832 ± 0.145 0.657 ± 0.180 0.701 ± 0.121 0.490 ± 0.151 TZ 0.716 ± 0.096 0.981 ± 0.059 0.601 ± 0.121 0.701 ± 0.087 0.582 ± 0.106 WG 0.761 ± 0.121 0.688 ± 0.190 0.893 ± 0.164 0.835 ± 0.108 0.581 ± 0.165 Region integration 0.858 ± 0.081 0.887 ± 0.157 0.825 ± 0.146 0.831 ± 0.080 0.712 ± 0.139 Open in a new tab Among the single-regional models, the TZ region performed relatively well, particularly in terms of SEN. For example, with the LR classifier, the TZ region achieved a SEN of 0.978, indicating strong recognition ability for positive samples. However, its SPE remained relatively low at 0.607, leading to a higher rate of misclassification for negative samples. In comparison, the PZ region showed weaker performance across metrics, with AUC values consistently below 0.7. For instance, with the LR classifier, the PZ region achieved an AUC of 0.693, reflecting limited diagnostic capability. The region integration model effectively addressed the limitations of single-regional models by combining features from the PZ and TZ regions, leading to significant overall performance improvements. For example, with the LR classifier, the region integration model achieved an AUC of 0.8815, which was substantially higher than the 0.722 for the TZ region alone. The SEN reached 0.977, while the SPE improved to 0.781. Similarly, with the LDA classifier, the region integration model achieved an AUC of 0.869, outperforming the single-regional results. For the SVM classifier, the region integration model achieved an AUC of 0.858, with SPE improving to 0.825, demonstrating its advantages over single-regional models. The ROC curves in Fig. 3 b further validated the advantages of the region integration model. The LR classifier achieved the highest AUC of 0.881, with its ROC curve closest to the top-left corner, indicating the best performance. The AUCs for LDA and SVM were 0.869 and 0.858, respectively, and all three classifiers substantially outperformed the single-regional results under the region integration strategy. While the TZ or WG exhibited relatively strong performance in specific metrics when used independently, the region integration model achieved comprehensive improvements in AUC, SEN, and SPE by fusing complementary features from the PZ, TZ, and WG regions. This demonstrated the effectiveness and overall superiority of multi-regional feature integration in the diagnosis of csPCa. Quantification of Feature Contribution to Classification We analyzed the contributions of features to the multi-regional radiomics model. Figure 4 a–c respectively presents the top 10 most influential features and their contribution distributions in the PZ, TZ, and WG regions. Each row in the figure represents a feature, the horizontal axis indicates the SHAP value reflecting the feature’s impact on the model’s prediction, and the color of each sample point represents the deviation direction and magnitude of the feature value: red indicates a positive deviation from the mean, blue indicates a negative deviation, and purple indicates values close to the mean. Fig. 4. Open in a new tab Beeswarm plots of feature contribution for multi-regional radiomics model at three prostate regions, namely (a) PZ, (b) TZ, and (c) WG From Fig. 4 , it was evident that features extracted from DWI and ADC images played a dominant role in each single-regional radiomics model. It was further noteworthy that while most of the contributing features were texture-based, they were closely associated with gray-level variations. This aligned with the clinical approach where radiologists evaluated the presence of csPCa by analyzing high and low signal intensities on DWI and ADC images in mpMRI. Table 4 summarizes the number of features after feature selection and their contributions from the three MRI parameters, namely DWI, ADC, and T2 WI, in the TZ, PZ, and WG. There were more features selected from DWI and ADC than T2 WI, regardless of TZ, PZ, or WG. The SHAP value analysis indicated that DWI had the highest degrees of contribution at 53.28%, followed by ADC at 39.52%, and T2 WI at 7.20%. These data suggested that DWI played a critical role in the diagnostic performance of the radiomics model, especially in TZ, where its contribution was particularly notable. Table 4. Quantitative contribution of MRI features Number of the features DWI ADC T2 WI TZ 9 6 3 PZ 5 5 2 WG 7 6 1 Total parametric SHAP value 1.599 1.186 0.216 Contribution degree (%) 53.28 39.52 7.20 Open in a new tab Comparative Results In the comparison experiments, we compared our method with three methods from the literature, and the results are listed in Tables 5 and 6 for internal and external test sets individually. Below is a detailed analysis of the comparison between our method and the three approaches from the literature: Table 5. Comparative test performance of different literature methods on internal test sets. The optimal result is in bold AUC ACC SEN SPE YI Interpretable-DL [ 39 ] 0.851 ± 0.065 0.852 ± 0.057 0.877 ± 0.085 0.832 ± 0.102 0.683 ± 0.069 Radiomics-Doc * [ 40 ] 0.907 ± 0.057 0.898 ± 0.046 0.979 ± 0.065 0.827 ± 0.097 0.807 ± 0.087 Swin UNETR [ 41 ] 0.874 ± 0.058 0.723 ± 0.072 0.923 ± 0.110 0.773 ± 0.128 0.698 ± 0.097 Ours 0.903 ± 0.064 0.879 ± 0.053 0.990 ± 0.043 0.819 ± 0.091 0.809 ± 0.089 Open in a new tab *Radiomics-Doc is semi-automated, relying on doctors’ manual segmentation of prostate zones, while other methods are fully automated Table 6. Comparative test performance of different literature methods on external test sets. The optimal result is in bold AUC ACC SEN SPE YI Interpretable-DL [ 39 ] 0.848 ± 0.094 0.864 ± 0.098 0.808 ± 0.152 0.891 ± 0.146 0.663 ± 0.087 Radiomics-Doc * [ 40 ] 0.878 ± 0.065 0.796 ± 0.074 0.978 ± 0.061 0.775 ± 0.109 0.753 ± 0.098 Swin UNETR [ 41 ] 0.845 ± 0.071 0.729 ± 0.081 0.944 ± 0.094 0.741 ± 0.129 0.685 ± 0.107 Ours 0.881 ± 0.055 0.803 ± 0.070 0.977 ± 0.062 0.781 ± 0.099 0.758 ± 0.088 Open in a new tab We compared Kim et al.’s [ 39 ] interpretable deep learning method, which also used T2, ADC, and DWI for diagnosing csPCa. Here, we termed it the Interpretable-DL method. Interpretable-DL operates directly on the whole-gland mpMRI and therefore does not require any prostate-zone segmentation. Unlike our approach, Kim et al. did not incorporate radiomics feature extraction in their model, relying solely on deep learning training. As a result, their model primarily depends on the raw pixel information from the images and lacks additional radiomics feature extraction and optimization. In contrast, our method combines radiomics feature extraction with deep learning models, and as shown in the table, our performance metrics significantly outperform Kim et al.’s method on both the internal and external test datasets. We also compared Li et al.’s [ 40 ] method, which is a radiomics method employing prostatic zonal ground truth (GT) drawn by doctors, named as Radiomics_Doc. Since the method relies on doctor-delineated GT of prostate zones, it can more accurately mark the lesion areas, leading to the extraction of more representative features, which results in relatively better performance in terms of metrics. However, this method has a clear limitation: it depends on the doctor’s delineation, which is time-consuming and labor-intensive, and it is constrained by the doctor’s experience and skill, leading to issues with scalability and consistency in practical applications. Moreover, the subjectivity of doctors may introduce biases, affecting the accuracy of the annotations. Our method, on the other hand, does not rely on GT, and its performance is almost on par with Li et al.’s method, and in some cases, even exceeds it. In addition, we compared the Swin UNETR model proposed by Tang et al. [ 41 ], using it for segmentation and combining it with radiomics feature extraction. Like our method, Swin UNETR relies on automatic segmentation, but its transformer-only design emphasizes global context at the expense of fine local details and boundary precision. In contrast, our CCT-Unet model [ 25 ], which combines the advantages of CNN and transformer, is able to capture both local features and global context more precisely, leading to superior segmentation accuracy. As seen in Tables 5 and 6 , our performance, especially on the core evaluation metric AUC, is superior. Discussion In this study, we proposed a multi-regional multiparametric radiomics method for csPCa diagnosis that integrated deep learning segmentation technology with radiomics feature extraction, enabling a comprehensive analysis of prostate mpMRI. Through the experimental analysis of the internal and external datasets, it was observed that the region integration model outperformed single-regional models in terms of diagnostic performance. In the region integration model, all three classifiers, i.e., LDA, LR, and SVM, demonstrated higher AUC values than those of the single-regional models. This showed that the region integration strategy effectively integrated information from multiple regions, enhancing the model’s diagnostic performance. In addition, our method also shows excellent performance in comparison with other methods. The innovation of our study was mainly reflected in the proposal and utilization of the multi-region integration strategy. Traditional csPCa diagnosis methods often relied on the analysis of a single region. However, our study broke through the spatial limitations of a single-region model by integrating the features of the PZ, TZ, and WG and could more comprehensively capture the pathological characteristics of each region of the prostate, thereby improving the accuracy of diagnosis. In addition, the combination of deep learning segmentation and radiomics analysis enabled automated region segmentation and feature extraction, which not only improved work efficiency but also ensured the consistency and repeatability of the diagnostic process. Finally, the feature contribution analysis based on SHAP further enhanced the interpretability of the model, clarified the importance of different MRI parameters in the model prediction, and improved the clinical usability of the diagnostic model. These innovative technical means provided new insights for the accurate diagnosis of csPCa and more reliable support for clinical decision-making. The SHAP-based feature contribution analysis provided quantitative insights into the importance of different MRI parameters in the diagnostic model. According to Table 4 , regardless of the TZ, PZ, or WG, the DWI and ADC contributed more features in the classification model and consistently showed higher contributions compared to T2 WI. This demonstrated that in radiomics-based csPCa diagnosis, the DWI and ADC features played a more pivotal role in distinguishing csPCa lesions. This finding aligns with clinical practice, where the DWI and ADC images are widely utilized to assess the aggressiveness of PCa. We observed that the TZ region consistently outperformed the PZ and WG regions across all classifiers (LDA, LR, and SVM) in terms of diagnostic performance. This finding was consistent with that of Esengur et al. [ 42 ], who reported that the TZ region achieved the most stable and accurate results in csPCa diagnosis. This advantage may stem from the TZ region’s clearer anatomical structure and more stable imaging characteristics, which enable the model to extract consistent and discriminative features. Additionally, benign and malignant lesions in the TZ often exhibit more distinct imaging patterns, further enhancing the model’s ability to differentiate them. Therefore, we believe the superior diagnostic performance of the TZ region is the result of multiple contributing factors, including structural regularity, imaging clarity, and a stronger correlation between clinical presentation and imaging features. In our study, we did not artificially balance the sample ratio between csPCa and non-csPCa cases but instead preserved the natural distribution observed in real-world clinical settings. In our study, the proportions of csPCa cases were 32.5%, 26.1%, and 30.5% in the internal, external, and whole datasets, respectively. These proportions align well with the reported csPCa detection rates (roughly 20–40%) in several high-quality studies [ 43 – 46 ]. We believe that maintaining the natural distribution allows for a more realistic evaluation of model performance in clinical practice, avoiding overfitting or optimistic bias toward the minority class. Our study still faces some challenges, such as the complexity of the feature extraction and selection process. Future research will focus on exploring more efficient feature extraction and selection strategies, such as end-to-end deep learning frameworks, to further improve the automation level of the model. In addition, a more precise segmentation model for both PZ and TZ could be developed in the future to further improve the diagnostic accuracy and robustness. The adaptability and stability of the model under various devices and imaging conditions could be ensured through multi-center dataset validation. Moreover, to address class imbalance, we plan to explore strategies such as cost-sensitive learning, minority-class augmentation, and threshold calibration, which will further enhance the model’s robustness under imbalanced conditions. Conclusion In this work, we have proposed a csPCa diagnostic approach that leverages deep learning and multi-regional multiparametric radiomics. Our approach ensures efficient, accurate, and non-invasive csPCa diagnosis through automatic segmentation of mpMRI images and comprehensive radiomics analysis. Our results highlight that the multi-regional multiparametric model substantially outperforms single-regional models, demonstrating clear advantages in diagnostic performance with notable improvements across multiple key metrics on both internal and external datasets. The feature contribution analysis reveals that the DWI and ADC features are crucial for csPCa diagnosis, aligning with clinical practice and reinforcing the model’s transparency and interpretability. Compared with existing methods, our proposed method demonstrates superior performance, underscoring its advantages. By integrating information from multiple prostatic regions and imaging parameters, our approach not only enhances diagnostic performance but also provides a solid foundation for future precise management of csPCa. Further research should focus on enhancing the model’s generalization capabilities, facilitating its broader clinical application and ultimately optimizing patient diagnosis and treatment outcomes. Author Contribution Xijun Liu: investigation, methodology, software, and writing—original draft. Rongzong Liu: validation, formal analysis, investigation and data curation. Haihao He: methodology, visualization, and software. Yifei Yan: methodology and software. Limin Zhang: conceptualization, validation, data curation, writing—review and editing, supervision, and funding acquisition. Qi Zhang: conceptualization, validation, resources, writing—review and editing, supervision, project administration, and funding acquisition. Funding This work was funded by the National Natural Science Foundation of China (no. 62071285), the Shanghai Municipal Health Commission (no. 202040439), the Eastern Scholars Program from Shanghai Municipal Education Commission, and supported by the Shanghai Technical Service Center of Science and Engineering Computing, Shanghai University. Data Availability The authors do not have permission to share data. Declarations Ethics Approval This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Medical Research Ethics Committee of Huashan Hospital (KY2011-009). Consent to Participate Informed consent was obtained from all individual participants included in the study. Consent for Publication All participants in this study gave their consent for publication. Competing interests The authors declare no competing interests. Footnotes Xijun Liu and Rongzong Liu contributed equally to this work and are co-first authors. Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Limin Zhang, Email: zhalim@fudan.edu.cn. Qi Zhang, Email: zhangq@t.shu.edu.cn. References 1. F. Bray, M. Laversanne, H. Sung, J. Ferlay, R.L. Siegel, I. Soerjomataram and A. Jemal Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: A Cancer Journal for Clinicians 2024;74:229–63. [ DOI ] [ PubMed ] 2. R.L. Siegel, A.N. Giaquinto and A. Jemal Cancer statistics, 2024. CA Cancer J Clin 2024;74:12–49. [ DOI ] [ PubMed ] 3. J. Pylväläinen, J. Hoffström, A. Kenttämies, A. Auvinen, T. Mirtti and A. Rannikko Risk of clinically significant prostate cancer after a nonsuspicious prostate MRI—a comparison with the general population. Cancer Epidemiology, Biomarkers & Prevention 2024;33:749–56. [ DOI ] [ PMC free article ] [ PubMed ] 4. A. Matoso and J.I. Epstein Defining clinically significant prostate cancer on the basis of pathological findings. Histopathology 2019;74:135–45. [ DOI ] [ PubMed ] 5. U. Stopka-Farooqui, V. Stavrinides, B.S. Simpson, H. Qureshi, L.M. Carmona Echevierra, H. Pye, et al. Combining tissue biomarkers with mpMRI to diagnose clinically significant prostate cancer. Analysis of 21 biomarkers in the PICTURE study. Prostate Cancer and Prostatic Diseases 2024; [ DOI ] [ PMC free article ] [ PubMed ] 6. J. Xiang, H. Yan, J. Li, X. Wang, H. Chen and X. Zheng Transperineal versus transrectal prostate biopsy in the diagnosis of prostate cancer: a systematic review and meta-analysis. World journal of surgical oncology 2019;17:1–11. [ DOI ] [ PMC free article ] [ PubMed ] 7. A. Matsukawa, T. Yanagisawa, K. Bekku, M.K. Parizi, E. Laukhtina, J. Klemm, et al. Comparing the performance of digital rectal examination and prostate-specific antigen as a screening test for prostate cancer: a systematic review and meta-analysis. European Urology Oncology 2024; [ DOI ] [ PubMed ] 8. W. Ji, Y. Wang, Z. Han, S. Wang and Y. Wang Construction of a clinically significant prostate cancer risk prediction model based on traditional diagnostic methods. Frontiers in Oncology 2024;14: [ DOI ] [ PMC free article ] [ PubMed ] 9. W.G. Mwende Advancements in Non-Invasive Imaging Techniques for Early Detection of Prostate Disorders: A Review. IAA Journal of Scientific Research 2024;11:14–18. 10. E.T.d.O. Correia, A. Baydoun, Q. Li, D.N. Costa and L.K. Bittencourt Emerging and anticipated innovations in prostate cancer MRI and their impact on patient care. Abdom Radiol 2024;49:3696–710. [ DOI ] [ PMC free article ] [ PubMed ] 11. G. Lee, A. Chatterjee, C. Harmath, I. Karademir, R. Engelmann, A. Yousuf, et al. Improving reader accuracy and specificity with the addition of hybrid multidimensional-MRI to multiparametric-MRI in diagnosing clinically significant prostate cancers. Abdominal Radiology 2023;48:3216-28. [ DOI ] [ PubMed ] [ Google Scholar ] 12. M.D. Greer, A.M. Brown, J.H. Shih, R.M. Summers, J. Marko, Y.M. Law, et al. Accuracy and agreement of PIRADSv2 for prostate cancer mpMRI: a multireader study. Journal of magnetic resonance imaging 2017;45:579-85. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. A. Chatterjee, C. Harmath and A. Oto New prostate MRI techniques and sequences. Abdominal Radiology 2020;45:4052–62. [ DOI ] [ PubMed ] 14. J.M. Castillo T, M. Arif, W.J. Niessen, I.G. Schoots and J.F. Veenland Automated classification of significant prostate cancer on MRI: a systematic review on the performance of machine learning applications. Cancers 2020;12:1606. [ DOI ] [ PMC free article ] [ PubMed ] 15. N. Desouza, S. Riches, N. Vanas, V. Morgan, S. Ashley, C. Fisher, et al. Diffusion-weighted magnetic resonance imaging: a potential non-invasive marker of tumour aggressiveness in localized prostate cancer. Clinical radiology 2008;63:774-82. [ DOI ] [ PubMed ] [ Google Scholar ] 16. P. Hambarde, S. Talbar, A. Mahajan, S. Chavan, M. Thakur and N. Sable Prostate lesion segmentation in MR images using radiomics based deeply supervised U-Net. Biocybernetics and Biomedical Engineering 2020;40:1421–35. 17. A. Ali, A. Du Feu, P. Oliveira, A. Choudhury, R.G. Bristow and E. Baena Prostate zones and cancer: lost in transition? Nat Rev Urol 2022;19:101–15. [ DOI ] [ PubMed ] 18. C. Tsuruta, K. Hirata, K. Kudo, N. Masumori and M. Hatakenaka DWI-related texture analysis for prostate cancer: Differences in correlation with histological aggressiveness and data repeatability between peripheral and transition zones. European radiology experimental 2022;6:1–12. [ DOI ] [ PMC free article ] [ PubMed ] 19. P. Patel, S. Nallandhighal, D. Scoville, B. Cotta, Z. Knuth, D. Triner, et al. Spatial Transcriptomic Profiling to Characterize the Nature of Peripheral-Versus Transition-zone Prostate Cancer. European Urology Focus 2024; [ DOI ] [ PubMed ] 20. E. Mylona, D.I. Zaridis, C.Ν. Kalantzopoulos, N.S. Tachos, D. Regge, N. Papanikolaou, et al. Optimizing radiomics for prostate cancer diagnosis: feature selection strategies, machine learning classifiers, and MRI sequences. Insights into Imaging 2024;15:265. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. B. Zheng, F. Mo, X. Shi, W. Li, Q. Shen, L. Zhang, et al. An Automatic Deep-Radiomics Framework for Prostate Cancer Diagnosis and Stratification in Patients with Serum Prostate-Specific Antigen of 4.0–10.0 ng/mL: A Multicenter Retrospective Study. Academic Radiology 2025; [ DOI ] [ PubMed ] 22. M. Hammad, M. Bakrey, A. Bakhiet, R. Tadeusiewicz, A.A.A. El-Latif and P. Pławiak A novel end-to-end deep learning approach for cancer detection based on microscopic medical images. Biocybernetics and Biomedical Engineering 2022;42:737–48. 23. A. Saha, M. Hosseinzadeh and H. Huisman End-to-end prostate cancer detection in bpMRI via 3D CNNs: effects of attention mechanisms, clinical priori and decoupled false positive reduction. Medical image analysis 2021;73:102155. [ DOI ] [ PubMed ] 24. N. Netzer, C. Eith, O. Bethge, T. Hielscher, C. Schwab, A. Stenzinger, et al. Application of a validated prostate MRI deep learning system to independent same-vendor multi-institutional data: demonstration of transferability. European Radiology 2023;33:7463-76. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Y. Yan, R. Liu, H. Chen, L. Zhang and Q. Zhang CCT-Unet: A U-shaped Network based on Convolution Coupled Transformer for Segmentation of Peripheral and Transition Zones in Prostate MRI. JBHI 2023;PP:1–11. [ DOI ] [ PubMed ] 26. A. Raj, A. Gass, P. Eisele, A. Dabringhaus, M. Kraemer and F.G. Zöllner A generalizable deep voxel-guided morphometry algorithm for the detection of subtle lesion dynamics in multiple sclerosis. Frontiers in Neuroscience 2024;18:1326108. [ DOI ] [ PMC free article ] [ PubMed ] 27. S. Albert, B.D. Wichtmann, W. Zhao, A. Maurer, J. Hesser, U.I. Attenberger, et al. Comparison of image normalization methods for multi-site deep learning. Applied Sciences 2023;13:8923. [ Google Scholar ] 28. S. Liqin, S. Dinggang and Q. Feihu Edge detection on real time using LOG filter. Proceedings of ICSIPNN'94. International Conference on Speech, Image Processing and Neural Networks 1994;37–40. 29. C.S. Burrus, R.A. Gopinath and H. Guo Wavelets and wavelet transforms. rice university, houston edition 1998;98: 30. H. Zou The Adaptive Lasso and Its Oracle Properties. Journal of the American Statistical Association 2006;101:1418–29. 31. Y. Liu, Z. Liu, X. Luo and H. Zhao Diagnosis of Parkinson's disease based on SHAP value feature selection. Biocybernetics and Biomedical Engineering 2022;42:856–69. 32. G. Hughes Youden's Index and the Weight of Evidence. Methods Inf Med 2015;54:198–99. [ DOI ] [ PubMed ] 33. S. Lundberg and S.-I. Lee A Unified Approach to Interpreting Model Predictions. 2017; 34. H.M. Balaha, A.E.-S. Hassan, R.A. Ahmed and M.H. Balaha Advancing eye disease detection: A comprehensive study on computer-aided diagnosis with vision transformers and SHAP explainability techniques. Biocybernetics and Biomedical Engineering 2025;45:23–33. 35. L.S. Shapley and A.E. Roth The Shapley value : essays in honor of Lloyd S. Shapley. 1988; 36. M.D.E. Goodyear, K. Krleza-Jeric and T. Lemmens The Declaration of Helsinki. BMJ 2007;335:624–25. [ DOI ] [ PMC free article ] [ PubMed ] 37. H. Huisman, P. Vos, G. Litjens, T. Hambrock, J. Barentsz, P. Abolmaesumi, et al. Computer Aided Detection of Prostate Cancer Using T2, DWI and DCE MRI: Methods and Clinical Applications. 2010;6367:4-14. [ Google Scholar ] 38. Y. Liu, K. Sung, G. Yang, S. Afshari Mirak, M. Hosseiny, A. Azadikhah, et al. Automatic Prostate Zonal Segmentation Using Fully Convolutional Network With Feature Pyramid Attention. IEEE Access 2019;7:163626–32. 39. H. Kim, D.J. Margolis, H. Nagar and M.R. Sabuncu Pulse sequence dependence of a simple and interpretable deep learning method for detection of clinically significant prostate cancer using multiparametric MRI. Academic Radiology 2023;30:966–70. [ DOI ] [ PubMed ] 40. C. Li, Z. Jin, C. Wei, G. Dai, J. Tu and J. Shen Comparison in prostate cancer diagnosis with PSA 4–10 ng/mL: radiomics-based model VS. PI-RADS v2. 1. BMC urology 2024;24:233. [ DOI ] [ PMC free article ] [ PubMed ] 41. Y. Tang, D. Yang, W. Li, H.R. Roth, B. Landman, D. Xu, et al. Self-supervised pre-training of swin transformers for 3d medical image analysis. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition 2022;20730–40. 42. O.T. Esengur, E. Stevenson, H. Stecko, N.S. Lay, D. Yang, J. Tetreault, et al. Assessing the Impact of Transition and Peripheral Zone PSA Densities Over Whole‐Gland PSA Density for Prostate Cancer Detection on Multiparametric MRI. The Prostate 2025;85:612-24. [ DOI ] [ PubMed ] [ Google Scholar ] 43. V. Kasivisvanathan, A.S. Rannikko, M. Borghi, V. Panebianco, L.A. Mynderse, M.H. Vaarala, et al. MRI-targeted or standard biopsy for prostate-cancer diagnosis. New England Journal of Medicine 2018;378:1767-77. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. M. Valerio, I. Donaldson, M. Emberton, B. Ehdaie, B.A. Hadaschik, L.S. Marks, et al. Detection of clinically significant prostate cancer using magnetic resonance imaging–ultrasound fusion targeted biopsy: a systematic review. European urology 2015;68:8-19. [ DOI ] [ PubMed ] [ Google Scholar ] 45. M. Eklund, F. Jäderling, A. Discacciati, M. Bergman, M. Annerstedt, M. Aly, et al. MRI-targeted or standard biopsy in prostate cancer screening. New England Journal of Medicine 2021;385:908-20. [ DOI ] [ PubMed ] [ Google Scholar ] 46. H.U. Ahmed, A.E.-S. Bosaily, L.C. Brown, R. Gabe, R. Kaplan, M.K. Parmar, et al. Diagnostic accuracy of multi-parametric MRI and TRUS biopsy in prostate cancer (PROMIS): a paired validating confirmatory study. The Lancet 2017;389:815-22. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The authors do not have permission to share data. Articles from Journal of Imaging Informatics in Medicine are provided here courtesy of Springer ACTIONS View on publisher site PDF (2.4 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Deep learning applied to cancer prostate histology slides : diagnosis, prediction of bio-chemical recurrence and genetic mutations#179409
• Deep Learning Signature for Predicting Occult Nodal Metastasis of Clinical N0 Lung Cancer#1047191
• Graphical machine learning for precision medicine: radiomic analysis of hepatocellular carcinoma. (Machine learning graphique pour la médecine de précision: analyse radiomique du carcinome hépatocellulaire).#489482
• Detection of ISUP≥2 Prostate Cancers Using Multiparametric MRI: Prospective Multicenter Comparison of the PI-RADS Score and an Artificial Intelligence System#902630
• Prospective Validation of an AI Model for Predicting Liver Metastasis in Colorectal Cancer#857883