arXiv:2604.13695v1 [cs.CV] 15 Apr 2026
Med-CAM: Minimal Evidence for Explaining Medical Decision Making Pirzada Suhail IIT Bombay
Aditya Anand IIT Bombay
Amit Sethi IIT Bombay
Abstract Reliable and interpretable decision-making is essential in medical imaging, where diagnostic outcomes directly influence patient care. Despite advances in deep learning, most medical AI systems operate as opaque black boxes, providing little insight into why a particular diagnosis was reached. In this paper, we introduce MedCAM, a framework for generating minimal and sharp maps as evidence-based explanations for Medical decision making via Classifier Activation Matching. Med-CAM trains a segmentation network from scratch to produce a mask that highlights the minimal evidence critical to model’s decision for any seen or unseen image. This ensures that the explanation is both faithful to the network’s behavior and interpretable to clinicians. Experiments show, unlike prior spatial explanation methods, such as Grad-CAM and attention maps, which yield only fuzzy regions of relative importance, Med-CAM with its superior spatial awareness to shapes, textures, and boundaries, delivers conclusive, evidence-based explanations that faithfully replicate the model’s prediction for any given image. By explicitly constraining explanations to be compact, consistent with model activations, and diagnostic alignment, MedCAM advances transparent AI to foster clinician understanding and trust in high-stakes medical applications such as pathology and radiology.
1. Introduction Explanations are increasingly being recognized as essential for understanding and trusting the decision-making of machine learning systems deployed in medical settings. In clinical workflows, where algorithmic predictions can directly influence diagnosis or treatment, it is not enough for a model to be accurate—it must also be interpretable and verifiable. Deep neural networks, despite their remarkable success across medical imaging tasks such as pathology, radiology, and dermatology, often operate as opaque black boxes. Their predictions arise from complex, highdimensional computations that are inaccessible to human
reasoning, leaving clinicians uncertain about whether the model’s conclusions stem from meaningful pathology or from confounding artifacts. In this context, minimal evidence becomes a key desideratum for trustworthy medical explanations. By isolating the smallest set of image regions sufficient to support a model’s diagnostic decision, one can reveal the true basis of the model’s reasoning in a form that is both humanunderstandable and clinically verifiable. Minimal explanations reveal the essential visual cues such as atypical cellular morphologies in histopathology or subtle opacities in radiographs, that directly drive the network’s prediction. Such evidence-centered explanations not only enhance clinical interpretability but also serve as a diagnostic sanity check, allowing practitioners to verify weather that the model attends to medically relevant structures or not. To address this need, we propose Med-CAM, a Medical Classifier Activation Matching framework for generating minimal, evidence-based explanations for medicinal decision making. Med-CAM learns a lightweight U-Net from scratch that produces a binary mask highlighting the minimal evidence necessary for a model’s output. The U-Net is trained individually on a single image within seconds, optimizing for consistency between the classifier’s activations on the original image and on the masked explanation. This activation-matching principle ensures that the explanation preserves both the internal reasoning and the diagnostic prediction of the underlying model. Compared to existing saliency-based approaches such as Grad-CAM and attention maps, Med-CAM offers several key advantages. Grad-CAM provides gradient-weighted heatmaps that highlight regions of relative importance but fail to capture precise spatial characteristics such as boundaries, textures, or surface continuity. In contrast, Med-CAM explicitly enforces activation and decision consistency, resulting in spatially aware, compact, and conclusive evidence maps that faithfully replicate the model’s diagnostic behavior. By grounding explanations in minimal, activationconsistent evidence, Med-CAM bridges the gap between model transparency and clinical interpretability, contributing toward the development of trustworthy medical AI.
2. Prior Work 2.1. Inversion Inversion methods aim to reconstruct inputs that elicit specific outputs or internal activations of a neural network seeking to visualize what a model has learned rather than why it makes a particular decision. Early work reconstructed representative patterns from multilayer perceptrons through gradient-based optimization, though such visualizations were often noisy or adversarial-like [13, 14, 26]. Subsequent studies explored evolutionary search and constrained optimization [38], followed by the introduction of prior-based regularization techniques such as smoothness constraints and pretrained generative models to enhance realism and interpretability [19–22, 39]. More recent methods stabilize the inversion process through learned surrogate landscapes [16] or reframe it within logical reasoning frameworks for deterministic reconstruction [33] while others [34] use generative modeling for inversion.
2.2. Explainability While inversion captures model behavior in aggregate, explainability focuses on generating faithful rationales for individual predictions. Explainable AI (XAI) has therefore emerged as a major research area [1, 8, 11], driven by the need to enhance trust, transparency, and accountability in high-stakes applications. Post-hoc attribution methods remain the dominant paradigm. LIME builds local surrogate models to approximate black-box predictors [9], while Anchors [24] extends this idea through high-precision, rule-based local explanations that define sufficient conditions for predictions. Grad-CAM highlights salient image regions by gradient-weighted class activation mapping [28], whereas DeepLIFT [30] backpropagates contribution scores relative to reference activations. The SHAP framework [18] unifies several additive feature-attribution approaches using Shapley values to assign consistent feature importance scores. Beyond local attributions, concept-based and structural methods attempt to align neural activations with semantically meaningful parts [15]. Architectural approaches [3] promote explicit weight–input alignment for interpretable transformations, while [32] introduces an auxiliary explainer network that produces class-specific binary masks identifying discriminative evidence regions. Parallel to these efforts, logic- and reasoning-based frameworks explore explanation as a formal process. [5] introduced theoretical foundations for necessary, sufficient, and complete reasons behind classifier decisions. Similarly, [29] proposed a symbolic approach to explain Bayesian network classifiers, identifying minimal sufficient feature sets responsible for classification. Abductive reasoning techniques [12] extend this logic-driven line of work by comput-
ing subset- or cardinality-minimal explanations with formal guarantees of faithfulness. Recent surveys [40] emphasize the need for quantitative evaluation criteria—combining fidelity, stability, and human-centered assessment—to move beyond purely visual interpretability. Interactive frameworks like [35] further demonstrate how explanations can guide human-in-the-loop model refinement.
2.3. Explainability in Medical Diagnostics In medical imaging, interpretability is not merely desirable—it is essential. [31] surveyed explainable deep learning methods for medical image analysis and highlighted major barriers to clinical adoption, including lack of interpretive rigor and standardized evaluation. [37] proposed a comprehensive taxonomy of explainable medical imaging methods categorized by anatomical region, modality, and transparency level. More recently, [4] emphasized the joint role of explainability and generalizability, showing that integrating CNNs with XAI methods such as GradCAM, XGradCAM, and LayerCAM enhances visual interpretability and diagnostic consistency across multiple medical datasets. [7] presents a focused scoping review on explainability in breast cancer detection identifying SHAP as the most widely used and clinically interpretable method, particularly suited to feature-level analysis and biomarker attribution in survival and risk prediction. Despite the steady progress of XAI in medicine, most existing frameworks emphasize saliency visualization or heuristic attribution without enforcing minimality—the principle that an explanation should contain only the smallest subset of evidence sufficient to preserve a model’s decision. Our proposed work Med-CAM, addresses this limitation by formulating explanation generation as an activationmatching inverse problem: we seek the minimal pre-image that preserves the classifier’s internal activations and decision, thereby producing compact, faithful, and diagnostically meaningful evidence maps.
3. Methodology Per-image explanation generation can be viewed as an inverse problem, aiming to recover a decision-wise conclusive and critical pre-image—a minimal subset of the input—that is sufficient to sustain a model’s prediction. While the region outside this subset acts as a distractor, containing information that does not contribute meaningfully to the decision and may even obscure it. Med-CAM aims to generate minimal and faithful evidence maps that explain the diagnostic decisions of a frozen medical classifier f on any given image x. To achieve this, we train a lightweight U-Net that outputs a binary mask m, where the masked explanation is defined as e = m ⊙ x, and non-critical regions are suppressed as shown in Figure
Figure 1. Overview of the Med-CAM framework. Given an input medical image x, a lightweight U-Net produces a binary mask m, yielding the masked explanation e = m ⊙ x. The mask highlights the minimal diagnostic evidence required to preserve the model’s decision.
1. The autoencoder is optimized individually for each image, requiring only a few seconds to converge, thus adapting dynamically to case-specific diagnostic cues. Training is guided by a composite loss function that integrates activation matching, output fidelity, and mask priors for minimality, along with a robustness constraint ensuring clinical sufficiency and stability.
3.1. Weighted Activation Matching The core principle of Med-CAM is that the explanation must preserve the classifier’s decision process at both the output and feature levels. This is enforced through activation alignment, distributional consistency and label match between the original image x and the masked explanation e. Activation Matching Loss: To ensure that the explanation elicits the same internal responses as the original input, we minimize X Lact = αℓ d ϕℓ (x), ϕℓ (e) , ℓ
where ϕℓ (·) denotes the post-ReLU activations at layer ℓ, d is a distance metric (mean squared error or cosine distance), and αℓ controls per-layer weighting. This loss ensures that
e reproduces the hierarchical activations responsible for diagnostic reasoning—such as tissue texture patterns, lesion morphology, or intensity gradients. KL Divergence Loss: To preserve the classifier’s probabilistic output distribution, we minimize the Kullback–Leibler divergence between the softmax outputs of the original and masked images: LKL = DKL softmax(f (x)) ∥ softmax(f (e)) . This encourages the masked explanation to produce the same confidence distribution as the original image, ensuring diagnostic consistency. Cross-Entropy Loss: To explicitly retain the predicted class label, we minimize LCE = − log pf (e) (y), where y is the top-1 class predicted by f (x). This enforces decision-level faithfulness: the masked explanation must independently yield the same diagnosis.
3.2. Mask Priors for Minimal Evidence To generate interpretable evidence maps that highlight diagnostically relevant regions, we impose priors on the mask m.
Figure 2. Med-CAM explanations on BACH H&E breast cancer histopathology slides using a ViT-16 classifier.
Area Loss for Sparsity: Larea = ∥m∥1 . This term penalizes large active regions, encouraging MedCAM to use the fewest possible pixels necessary for preserving the model’s decision corresponding to the principle of minimality. Binarization Loss for Crispness: 2
Lbin = ∥m − m ∥1 . This drives the mask toward binary values (0 or 1), producing clear, interpretable evidence maps rather than diffuse saliency heatmaps. Total Variation Loss for Smoothness: X Ltv = |mi,j − mi+1,j | + |mi,j − mi,j+1 | . i,j
This regularizes the spatial structure of the mask, promoting contiguous evidence regions and suppressing isolated activations critical for delineating anatomical boundaries.
3.3. Abductive Robustness Constraint Minimality alone does not ensure that the explanation is clinically sufficient. To enforce robustness, we introduce an abductive constraint ensuring that randomizing irrelevant regions does not alter the model’s decision. Given a perturbed background r, we define ẽ = m ⊙ x + (1 − m) ⊙ r,
and require that f (ẽ) yield the same label as f (x) via the loss Lrob = − log pf (ẽ) (y). This ensures that explanations are faithful even when nonevidence regions are replaced with random or domainshifted content—an important criterion for robustness in medical interpretation.
3.4. Training Objective The Med-CAM U-Net is optimized using a composite loss LEXP that enforces activation alignment, minimality, and robustness. For clarity, we group the terms as: LAM = λact Lact + λCE LCE + λKL LKL , LMIN = λarea Larea + λbin Lbin + λtv Ltv , LROB = λrob Lrob . The complete objective is: LEXP = LAM + LMIN + LROB . With appropriate coefficient choices {λ· }, Med-CAM learns binary, sparse, and spatially coherent masks that retain the classifier’s decision and internal activations while suppressing irrelevant context. Unlike Grad-CAM or attention maps, which provide coarse regions of relative importance, Med-CAM produces conclusive pixel-level evidence maps that isolate the minimal diagnostic cues essential to the model’s prediction.
Figure 3. Med-CAM explanations on brain tumor MRI (left) and IDRiD retinal fundus images (right).
4. Results Med-CAM is model-agnostic and operates on any frozen classifier and any seen or unseen image at inference time. Because the U-Net [25] explainer is trained per-image in a few seconds, the framework naturally adapts to casespecific visual cues without requiring dataset-level retraining. We evaluate Med-CAM on four diverse medical imaging modalities—dermatology, histopathology, MRI, and retinal fundus—using four architectures: ViT-16 [6] on BACH [2](89% test accuracy), ConvNeXt-Small [17] on HAM10000 [36](85%), ResNet-18 [10] on IDRiD [23] (79%), and MobileNet-V2 [27] on Brain Tumor MRI (82%). Figure 2 shows an image from each BACH class (Normal, Benign, In Situ, Invasive) with corresponding MedCAM explanations. Histology images exhibit highly irregular, multi-scale patterns, making explanation generation particularly challenging. Med-CAM successfully adapts to these diverse morphologies, highlighting luminal structures, glandular boundaries, and malignant epithelial clusters. Using relative weights of 10:100:10 for activation matching, minimality, and robustness, we observe that Med-CAM consistently isolates concise diagnostic evidence improving classifier’s confidence from an average of 85% on the original images to 96% on the Med-CAM explanations. In Figure 3 for MRI, we include one healthy and one tumor-containing slice. Med-CAM highlights the outer brain boundary and central parenchyma for normal cases, and selectively isolates only the tumor core and surround-
ing edema for pathological cases. For IDRiD, we show images from retinopathy grades 2 and 4. Grade 2 exhibits mild lesions, and Med-CAM correspondingly identifies a very small, compact set of pixels capturing microaneurysms and early exudates. Grade 4 contains large, clinically significant lesions highlighted by Med-CAM. Figure 4 shows results for four HAM10000 classes (Melanoma, Nevus, Vascular Lesion, Dermatofibroma) in which Med-CAM produces explanations that almost resemble segmentation masks–highlighting pigment networks, lesion borders, vascular blobs, and firm nodular structures with remarkable precision. With relative loss weights of 10:150:20, the evidence masks remain compact but highly informative, increasing classification confidence from an average of 75% to 93% after masking.
5. Comparisons We compare Med-CAM against the widely used GradCAM on the HAM10000 dataset (Figure 4). Grad-CAM heatmaps highlight large, smooth regions of relative importance but lack spatial awareness of fine lesion structure. Grad-CAM explanations fragment into multiple disconnected blobs despite the lesion being compact. Their activations frequently diffuse into background skin, illumination artifacts, vellus hair, and regions not diagnostic of the underlying condition. Moreover, Med-CAM, in contrast, produces crisp, binary explanations with sharply defined boundaries that adhere to the exact lesion morphology. The method precisely captures ir-
Figure 4. Med-CAM vs Grad-CAM on HAM10000.
regular lesion outlines, pigment asymmetry, and subtle textural cues essential for melanoma and nevus differentiation. Unlike Grad-CAM, Med-CAM’s evidence masks are minimal yet guaranteed to preserve the classifier’s original diagnosis. Empirically, Grad-CAM often spreads its attention across both lesion and non-lesion regions, while Med-CAM isolates only the discriminative pixels, returning a compact pre-image that is decision-equivalent to the full input.
6. Conclusion We introduced Med-CAM, a framework to generate minimal evidential explanations for medical image classifiers. Med-CAM consistently isolates the smallest diagnostically sufficient regions while preserving the model’s decision. By offering precise, compact, and clinically verifiable evidence maps, Med-CAM advances the reliability and interpretability of medical AI systems.
References [1] Sajid Ali, Tamer Abuhmed, Shaker El-Sappagh, Khan Muhammad, Jose M. Alonso-Moral, Roberto Confalonieri, Riccardo Guidotti, Javier Del Ser, Natalia Dı́az-Rodrı́guez, and Francisco Herrera. Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy artificial intelligence. Information Fusion, 99:101805, 2023. 2 [2] Guilherme Aresta, Teresa Araújo, Scotty Kwok, Sai Saketh Chennamsetty, Mohammed Safwan, Varghese Alex, Bahram Marami, Marcel Prastawa, Monica Chan, Michael Donovan, Gerardo Fernandez, Jack Zeineh, Matthias Kohl, Christoph Walz, Florian Ludwig, Stefan Braunewell, Maximilian Baust, Quoc Dang Vu, Minh Nguyen Nhat To, Eal Kim, Jin Tae Kwak, Sameh Galal, Veronica Sanchez-Freire, Nadia Brancati, Maria Frucci, Daniel Riccio, Yaqi Wang, Lingling Sun, Kaiqiang Ma, Jiannan Fang, Ismael Kone, Lahsen Boulmane, Aurélio Campilho, Catarina Eloy, António Polónia, and Paulo Aguiar. Bach: Grand challenge on breast cancer histology images. Medical Image Analysis, 56:122–139,
2019. 5 [3] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability, 2022. 2 [4] Ahmad Chaddad, Yan Hu, Yihang Wu, Binbin Wen, and Reem Kateb. Generalizable and explainable deep learning for medical image computing: An overview. Current Opinion in Biomedical Engineering, 33:100567, 2025. 2 [5] Adnan Darwiche and Auguste Hirth. On the reasons behind decisions, 2020. 2 [6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 5 [7] Amirehsan Ghasemi, Soheil Hashtarkhani, David Schwartz, and Arash Shaban-Nejad. Explainable artificial intelligence in breast cancer detection and risk prediction: A systematic scoping review, 2024. 2 [8] Leilani H. Gilpin, David Bau, Ben Z. Yuan, Ayesha Bajwa, Michael A. Specter, and Lalana Kagal. Explaining explanations: An overview of interpretability of machine learning. 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA), pages 80–89, 2018. 2 [9] Nicholas Hamilton, Adam Webb, Matt Wilder, Ben Hendrickson, Matt Blanck, Erin Nelson, Wiley Roemer, and Timothy C. Havens. Enhancing visualization and explainability of computer vision models with local interpretable model-agnostic explanations (lime). In 2022 IEEE Symposium Series on Computational Intelligence (SSCI), pages 604–611, 2022. 2 [10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 5 [11] Weiche Hsieh, Ziqian Bi, Chuanqi Jiang, Junyu Liu, Benji Peng, Sen Zhang, Xuanhe Pan, Jiawei Xu, Jinlang Wang, Keyu Chen, Pohsun Feng, Yizhu Wen, Xinyuan Song, Tianyang Wang, Ming Liu, Junjie Yang, Ming Li, Bowen Jing, Jintao Ren, Junhao Song, Hong-Ming Tseng, Yichao Zhang, Lawrence K. Q. Yan, Qian Niu, Silin Chen, Yunze Wang, and Chia Xin Liang. A comprehensive guide to explainable ai: From classical models to llms, 2024. 2 [12] Alexey Ignatiev, Nina Narodytska, and Joao Marques-Silva. Abduction-based explanations for machine learning models, 2018. 2 [13] C.A. Jensen, R.D. Reed, R.J. Marks, M.A. El-Sharkawi, Jae-Byung Jung, R.T. Miyamoto, G.M. Anderson, and C.J. Eggen. Inversion of feedforward neural networks: algorithms and applications. Proceedings of the IEEE, 87(9): 1536–1549, 1999. 2 [14] J Kindermann and A Linden. Inversion of neural networks by gradient descent. Parallel Computing, 14(3):277–286, 1990. 2 [15] Jae Hee Lee, Georgii Mikriukov, Gesina Schwalbe, Stefan Wermter, and Diedrich Wolter. Concept-based explanations in computer vision: Where are we and where could we go? In Computer Vision – ECCV 2024 Workshops, pages 266– 287, Cham, 2025. Springer Nature Switzerland. 2
[16] Ruoshi Liu, Chengzhi Mao, Purva Tendulkar, Hao Wang, and Carl Vondrick. Landscape learning for neural network inversion, 2022. 2 [17] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 5 [18] Scott Lundberg and Su-In Lee. A unified approach to interpreting model predictions, 2017. 2 [19] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them, 2014. 2 [20] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks, 2015. [21] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks, 2016. [22] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space, 2017. 2 [23] Prasanna Porwal, Samiksha Pachade, Ravi Kamble, Manesh Kokare, Girish Deshmukh, Vivek Sahasrabuddhe, and Fabrice Meriaudeau. Indian diabetic retinopathy image dataset (idrid): A database for diabetic retinopathy screening research. Data, 3(3), 2018. 5 [24] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Anchors: high-precision model-agnostic explanations. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2018. 2 [25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. 5 [26] Emad W. Saad and Donald C. Wunsch. Neural network explanation using inversion. Neural Networks, 20(1):78–93, 2007. 2 [27] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks, 2019. 5 [28] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 128(2):336–359, 2019. 2 [29] Andy Shih, Arthur Choi, and Adnan Darwiche. A symbolic approach to explaining bayesian network classifiers. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, page 5103–5111. AAAI Press, 2018. 2 [30] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences, 2019. 2 [31] Amitojdeep Singh, Sourya Sengupta, and Vasudevan Lakshminarayanan. Explainable deep learning models in medical image analysis, 2020. 2
[32] Steven Stalder, Nathanaël Perraudin, Radhakrishna Achanta, Fernando Perez-Cruz, and Michele Volpi. What you see is what you classify: Black box attributions, 2022. 2 [33] Pirzada Suhail. Network inversion of binarised neural nets. In The Second Tiny Papers Track at ICLR 2024, 2024. 2 [34] Pirzada Suhail and Amit Sethi. Network inversion of convolutional neural nets. In Muslims in ML Workshop co-located with NeurIPS 2024, 2024. 2 [35] Stefano Teso, Öznur Alkan, Wolfang Stammer, and Elizabeth Daly. Leveraging explanations in interactive machine learning: An overview, 2022. 2 [36] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data, 5(1), 2018. 5 [37] Bas H.M. van der Velden, Hugo J. Kuijf, Kenneth G.A. Gilhuijs, and Max A. Viergever. Explainable artificial intelligence (xai) in deep learning-based medical image analysis. Medical Image Analysis, 79:102470, 2022. 2 [38] Eric Wong. Neural network inversion beyond gradient descent. In WOML NIPS, 2017. 2 [39] Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization, 2015. 2 [40] Jianlong Zhou, Amir H. Gandomi, Fang Chen, and Andreas Holzinger. Evaluating the quality of machine learning explanations: A survey on methods and metrics. Electronics, 10 (5), 2021. 2