Interpretable Deep Learning Models for Breast Cancer Diagnosis Using Medical Imaging

Sami Alnawali; Muhammad Atif Saeed; Abdul Hadi; Azka Atiq; Akhtar Jamil; Ahmad Din

cs.CVOct 27, 2026

Interpretable Deep Learning Models for Breast Cancer Diagnosis Using Medical Imaging

Authors: Sami Alnawali, Muhammad Atif Saeed, Abdul Hadi, Azka Atiq, Akhtar Jamil, Ahmad Din

Organizations: Information Systems Department College of Computer Science and Information Technology King Faisal University, Alahsaa, Saudi Arabia · Department of Artificial Intelligence & Data Science National University of Computer and Emerging Sciences Islamabad, Pakistan · Department of Computer Science National University of Computer and Emerging Sciences Islamabad, Pakistan

Abstract

Deep learning shows promise for breast cancer diagnosis, but limited interpretability hinders clinical adoption. We benchmark ten ImageNet-pretrained models (ResNet50, VGG19, InceptionV3, DenseNet121, Xception, MobileNetV2, EfficientNet, ConvNeXt, ViT, and Swin Transformer) for benign-versus-malignant classification on the BreakHis histopathology dataset using a patient-level split, and explain their predictions with Grad-CAM, LIME, Integrated Gradients, and LRP. DenseNet121 performed best, with 99.4% test accuracy and an AUC-ROC of 0.999, and significantly outperformed all other models (p < 0.001). Transformer-based models trailed the CNNs. The explanations consistently highlighted nuclei-rich and glandular regions of diagnostic relevance. Combining transfer learning with explainable AI can thus yield accurate, interpretable tools for breast cancer diagnosis.

Figures & tables

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	89.2	90.7	95.5	93.0	70.3	0.936
Test (All Mag.)	99.4	99.9	99.5	99.6	99.2	0.999
Test (40 $\times$ )	100.0	100.0	100.0	100.0	100.0	1.000
Test (100 $\times$ )	100.0	100.0	100.0	100.0	100.0	1.000
Test (200 $\times$ )	100.0	100.0	100.0	100.0	100.0	1.000
Test (400 $\times$ )	97.5	99.4	97.7	98.5	96.4	0.998

Table 1: Performance of DenseNet-121 on the BreakHis dataset

Figure 1: Performance metrics comparison of DenseNet121 on the BreakHis dataset across validation and test sets at different magnifications (40X, 100X, 200X, 400X).

Panel of Figure 2 — Figure 2: Confusion matrices of the DenseNet-121 model across 40X, 100X, 200X, and 400X magnifications on the BreakHis dataset.

Panel of Figure 3 — Figure 3: Validation and test performance of the DenseNet-121 model on the BreakHis dataset across all magnifications.

Panel of Figure 4 — Figure 4: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h) for the DenseNet-121 model evaluated on the BreakHis dataset across 40X, 100X, 200X, and 400X magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	93.1	96.1	94.3	95.2	89.7	0.979
Test (All Mag.)	86.2	83.6	86.0	84.8	86.3	0.940
Test (40 $\times$ )	85.9	87.8	80.4	83.9	90.6	0.946
Test (100 $\times$ )	86.3	86.4	84.3	85.4	88.1	0.941
Test (200 $\times$ )	85.5	83.2	83.2	83.2	87.3	0.959
Test (400 $\times$ )	84.3	78.7	86.7	82.5	82.4	0.915

Table 2: Performance of EfficientNet on the BreakHis dataset

Figure 5: Model performance metrics on the BreakHis dataset across validation and test sets at different magnifications.

Panel of Figure 6 — Figure 6: Confusion matrices of the EfficientNet model across 40X, 100X, 200X, and 400X magnifications on the BreakHis dataset.

Panel of Figure 7 — Figure 7: Validation and test performance of the EfficientNet model on the BreakHis dataset across all magnifications.

Panel of Figure 8 — Figure 8: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h) for the EfficientNet model evaluated on the BreakHis dataset across 40X, 100X, 200X, and 400X magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	80.9	80.6	91.1	85.6	64.2	0.881
Test (All Mag.)	77.1	85.1	89.1	87.0	2.4	0.653
Test (40 $\times$ )	76.1	83.6	89.5	86.4	0.0	0.450
Test (100 $\times$ )	79.1	84.7	92.1	88.3	2.7	0.708
Test (200 $\times$ )	79.7	88.3	88.8	88.6	8.7	0.823
Test (400 $\times$ )	76.4	85.2	88.0	86.6	3.5	0.780

Table 3: Performance of ConvNeXt on the BreakHis dataset

Figure 9: Performance metrics of the ConvNeXt model across validation and test sets at multiple magnifications.

Panel of Figure 10 — Figure 10: Confusion matrices of the ConvNeXt model across 40X, 100X, 200X, and 400X magnifications on the BreakHis dataset.

Panel of Figure 11 — Figure 11: Validation and test performance of the ConvNeXt model on the BreakHis dataset across all magnifications.

Panel of Figure 12 — Figure 12: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h) for the ConvNeXt model evaluated on the BreakHis dataset across 40X, 100X, 200X, and 400X magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	94.3	96.2	95.9	96.1	89.9	0.980
Test (All Mag.)	84.6	79.5	88.4	83.7	81.5	0.933
Test (40 $\times$ )	87.2	86.0	86.0	86.0	88.2	0.937
Test (100 $\times$ )	81.2	77.0	85.9	81.3	76.9	0.924
Test (200 $\times$ )	85.1	80.4	86.3	83.3	84.1	0.938
Test (400 $\times$ )	86.9	78.3	95.9	86.2	80.2	0.943

Table 4: Performance of InceptionV3 on the BreakHis dataset

Figure 13: Performance of the Inception model shows high validation accuracy (94.5%, AUC = 0.98) and stable test results (81–87% accuracy, AUC = 0.93–0.95), with recall generally exceeding precision across magnifications.

Panel of Figure 14 — Figure 14: Confusion matrices of the Inception model across 40X, 100X, 200X, and 400X magnifications

Panel of Figure 15 — Figure 15: Validation and test performance of the Inception model on the BreakHis dataset across all magnifications.

Panel of Figure 16 — Figure 16: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h) for the Inception model evaluated on the BreakHis dataset across 40X, 100X, 200X, and 400X magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	91.8	95.9	92.7	94.3	89.5	0.975
Test (All Mag.)	84.7	83.5	81.9	82.7	86.9	0.933
Test (40 $\times$ )	81.6	84.0	73.8	78.6	88.2	0.923
Test (100 $\times$ )	84.7	82.0	86.8	84.3	82.8	0.928
Test (200 $\times$ )	86.4	85.0	83.2	84.0	88.9	0.956
Test (400 $\times$ )	84.3	81.6	81.6	81.6	86.3	0.927

Table 5: Performance of MobileNetV2 on the BreakHis dataset

Figure 17: MobileNet performance across validation and test sets at multiple magnifications

Panel of Figure 18 — Figure 18: Confusion matrices of the MobileNetV2 model across 40X, 100X, 200X, and 400X magnifications on the BreakHis dataset.

Panel of Figure 19 — Figure 19: Validation and test performance of the MobileNetV2 model

Panel of Figure 20 — Figure 20: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h) for the MobileNetV2 model evaluated on the BreakHis dataset across 40X, 100X, 200X, and 400X magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	93.2	96.2	94.4	95.3	90.2	0.978
Test (All Mag.)	83.6	79.1	86.2	82.5	81.5	0.919
Test (40 $\times$ )	87.2	87.4	84.1	85.7	89.8	0.914
Test (100 $\times$ )	82.4	80.6	82.6	81.6	82.1	0.925
Test (200 $\times$ )	83.3	76.4	88.4	81.9	79.4	0.928
Test (400 $\times$ )	80.8	73.3	86.7	79.4	76.3	0.907

Table 6: Performance of ResNet50 on the BreakHis dataset

Figure 21: Performance metrics of ResNet across validation and test sets at multiple magnifications.

Panel of Figure 22 — Figure 22: Confusion matrices of the ResNet model across 40X, 100X, 200X, and 400X magnifications

Panel of Figure 23 — Figure 23: Validation and test performance of the ResNet model on the BreakHis dataset across all magnifications.

Panel of Figure 24 — Figure 24: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h).

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Test (All Mag.)	79.5	73.6	84.6	78.7	75.3	0.860
Test (40 $\times$ )	85.0	81.6	86.9	84.2	83.5	0.929
Test (100 $\times$ )	79.2	74.6	85.1	79.5	73.9	0.874
Test (200 $\times$ )	77.8	69.8	85.3	76.8	72.2	0.834
Test (400 $\times$ )	77.7	70.1	83.7	76.3	73.3	0.821

Table 7: Performance of Swin Transformer Tiny on the BreakHis dataset

Figure 25: Performance metrics of the Swin Transformer Tiny across different magnifications

Figure 26: Training curves of Swin Transformer Tiny

Panel of Figure 27 — Figure 27: Confusion matrices of the Swin Transformer Tiny model across 40X, 100X, 200X, and 400X magnifications

Figure 28: Performance metrics of the VGG19 model across validation and test sets at multiple magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	93.1	95.9	94.6	95.2	89.3	0.975
Test (All Mag.)	81.3	76.5	84.1	80.1	79.0	0.905
Test (40 $\times$ )	80.3	75.6	84.1	79.6	77.2	0.899
Test (100 $\times$ )	78.8	74.5	84.3	79.1	73.9	0.914
Test (200 $\times$ )	84.2	76.8	90.5	83.1	79.4	0.938
Test (400 $\times$ )	80.4	74.8	81.6	78.1	79.4	0.892

Table 8: Performance of VGG19 on the BreakHis dataset

Panel of Figure 29 — Figure 29: Confusion matrices of the VGG19 model across 40X, 100X, 200X, and 400X magnifications on the BreakHis dataset.

Panel of Figure 30 — Figure 30: Validation and test performance of the VGG19 model on the BreakHis dataset across all magnifications.

Panel of Figure 31 — Figure 31: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h) for the VGG19 model evaluated on the BreakHis dataset across 40X, 100X, 200X, and 400X magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	86.5	90.9	90.5	90.7	75.9	0.892
Test (All Mag.)	77.6	71.1	84.3	77.2	72.2	0.858
Test (40 $\times$ )	79.1	71.0	91.6	80.0	68.5	0.856
Test (100 $\times$ )	72.6	68.4	78.5	73.1	67.2	0.789
Test (200 $\times$ )	81.0	73.9	86.3	79.6	77.0	0.893
Test (400 $\times$ )	79.0	71.9	83.7	77.4	75.6	0.862

Table 9: Performance of Vision Transformer (ViT) on the BreakHis dataset

Figure 32: ViT model performance across validation and test sets at multiple magnifications. Validation achieved strong results (Accuracy = 0.86, Precision = 0.91, Recall = 0.90, F1 = 0.91, AUC-ROC = 0.89).

Panel of Figure 33 — Figure 33: Confusion matrices of the Vision Transformer (ViT) model across 40X, 100X, 200X, and 400X magnifications on the BreakHis dataset.

Panel of Figure 34 — Figure 34: Validation and test performance of the Vision Transformer (ViT) model on the BreakHis dataset across all magnifications.

Panel of Figure 35 — Figure 35: Receiver Operating Characteristic (ROC) curves (a–d) and Precision–Recall (PR) curves (e–h) for the Vision Transformer (ViT) model evaluated on the BreakHis dataset across 40X, 100X, 200X, and 400X magnifications.

Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)	AUC
Validation (All Mag.)	94.5	95.7	96.9	96.3	88.3	0.985
Test (All Mag.)	85.8	80.0	91.2	85.2	81.5	0.939
Test (40 $\times$ )	87.2	82.9	90.7	86.6	84.3	0.945
Test (100 $\times$ )	84.3	80.9	87.6	84.1	81.3	0.935
Test (200 $\times$ )	87.3	80.2	93.7	86.4	82.5	0.961
Test (400 $\times$ )	83.8	74.0	95.9	83.6	74.8	0.928

Table 10: Performance of Xception on the BreakHis dataset

Figure 36: Performance metrics comparison of the Xception model across validation and test sets at multiple magnifications (40X, 100X, 200X, 400X).

Figure 37: Comparison of ten deep learning models on the test set (all magnifications) across accuracy, precision, recall, specificity, F1, and AUC-ROC.

Figure 38: Top-5 performing models across evaluation metrics on the test set (all magnifications).

Panel of Figure 39 — Figure 39: Explainable AI (XAI) visualizations of DenseNet121 predictions on the BreakHis dataset using (a) Grad-CAM, (b) Integrated Gradients, (c) LIME, and (d) Layer-wise Relevance Propagation (LRP).

Panel of Figure 40 — Figure 40: XAI visualizations of EfficientNet predictions on the BreakHis dataset using (a) Grad-CAM, (b) Integrated Gradients, (c) LIME, and (d) Saliency Maps.

Panel of Figure 41 — Figure 41: XAI visualizations of InceptionV3 predictions on the BreakHis dataset using (a) Grad-CAM, (b) Integrated Gradients, (c) LIME, and (d) Layer-wise Relevance Propagation (LRP).

Panel of Figure 42 — Figure 42: XAI visualizations of MobileNetV2 predictions on the BreakHis dataset using (a) Grad-CAM, (b) Integrated Gradients, (c) LIME, and (d) LRP.

Model	Arch.	Acc. (%)	Prec. (%)	Recall (%)	Spec. (%)	F1 (%)	AUC	$\bm{\kappa}$	MCC	F-Rank
DenseNet121	CNN	99.4 [98.6, 99.7]	99.9	99.5	99.2	99.6	0.999 [0.9963, 1.0000]	0.986	0.984	1.00
EfficientNet	CNN	86.2 [83.6, 88.4]	83.6	86.0	86.3	84.8	0.940 [0.9192, 0.9608]	0.685	0.700	3.62
Xception	CNN	85.8 [83.2, 88.1]	80.0	91.2	81.5	85.2	0.939 [0.9180, 0.9600]	0.676	0.689	3.50
InceptionV3	CNN	84.6 [81.9, 86.9]	79.5	88.4	81.5	83.7	0.933 [0.9111, 0.9549]	0.648	0.665	4.00
MobileNetV2	CNN	84.7 [82.0, 87.0]	83.5	81.9	86.9	82.7	0.933 [0.9111, 0.9549]	0.651	0.673	4.12
ResNet50	CNN	83.6 [80.9, 86.0]	79.1	86.2	81.5	82.5	0.919 [0.8950, 0.9430]	0.625	0.645	5.25

Table 11: Comprehensive statistical performance summary of all models on the BreakHis test set.

Figure 43: Statistical comparison of the 10 models tested in the patient-stratified BreakHis test set, with test accuracy and AUC-ROC and their 95% confidence intervals quantifying predictive uncertainty and comparative performance.

Figure 44: Friedman and Nemenyi non-parametric analyses comparing the ten architectures across four magnification levels that display statistically significant differences and average performance ranking of the evaluated models.

Figure 45: Magnification sensitivity analysis showing the robustness and stability of the model performance at different histopathological magnifications (40×, 100×, 200× and 400×) using the accuracy, AUC-ROC and deviation from accuracy metrics.

Explainable Medical Image Analysis

Explainable Medical Image Analysis develops methods that make medical image predictions understandable, verifiable, and useful for clinical review.

See all 256 papers

Architecture-aware Robustness Evaluation of Explainable Deep Learning for Breast Cancer Diagnosis
Sep 18, 2026Balenthira Thanusanth, Selvarajah Thuseethan, Roshan G. Ragel +3Explainable Medical Image Analysis Explainability Evaluation
University of Peradeniya, Sri Lanka · Charles Darwin University, Australia · Aneurin Bevan University Health Board, United Kingdom
Bridging visual saliency and large language models for explainable deep learning in medical imaging
May 7, 2026Paul Valery Nguezet, Elie Tagne Fute, Yusuf Brima +2Radiology Report Generation Medical Image Classification
Department of Mathematics and Computer science, University of Dschang, Dschang, 67, Cameroon. · Computer Vision, Institute of Cognitive Science, Osnabrück University, Osnabrueck, D-49090, Lower Saxony, Germany. · Department of Mathematics, Rhodes University, Grahamstown, Eastern Cape, 6140, Eastern Cape, South Africa. +1
Trusting Right Predictions for Wrong Reasons: A LIME Based Analysis of Deep Learning Interpretability in Lung Cancer Diagnosis
Jun 14, 2026Samarpan Poudel, Vladislav D VekslerExplainable Artificial Intelligence Explanation Stability
Caldwell University School of Business and Computer Science Caldwell, NJ 07006

Sep 18, 2026J/K move · Enter open · S save

Architecture-aware Robustness Evaluation of Explainable Deep Learning for Breast Cancer Diagnosis

Balenthira Thanusanth, Selvarajah Thuseethan, Roshan G. Ragel, Bimali S. Weerakoon+2

University of Peradeniya, Sri Lanka · Charles Darwin University, Australia · Aneurin Bevan University Health Board, United Kingdom

Explainable Medical Image Analysis Explainability Evaluation Breast Cancer Medical Image Classification

Open paper AskPDF arXiv

Explainable Artificial Intelligence (XAI) has become essential in medical image analysis to ensure transparency of deep learning (DL)-based diagnostic systems. However, selecting appropriate XAI techniques for breast cancer recognition remains largely ad hoc, with limited systematic evaluation across different DL architectures. This study presents a systematic architecture-aware evaluation protocol to assess the effectiveness of nine widely used XAI techniques across four categories of DL models: very deep, lightweight, transformer-based and hybrid neural networks. The evaluation is conducted on a breast ultrasound dataset comprising 780 images using clinically aligned spatial metrics, including Pointing Game, Intersection over Union and Mean Coverage, to quantify agreement between generated explanations and expert-annotated lesion regions. Results indicate that explanation quality is primarily influenced by the interaction between model architecture and XAI method, rather than any single technique consistently outperforming others. Hybrid architectures produce more spatially coherent explanations, while lightweight and transformer-based models exhibit greater variability across methods. The findings show that no single technique generalises across architectures and evaluation criteria, emphasising the need for joint selection of DL models and XAI techniques. Explainability depends on both model design and explanation strategy and should not be considered independently. This work provides a structured evaluation protocol and practical guidance for selecting XAI techniques in breast cancer diagnosis, supporting more transparent clinical decision-support systems. \textcolor{blue}{Code is publicly available at https://github.com/Nishan-Charlie/Explainable-AI}

Abstract

Explainable Medical Image Analysis

Explore similar work

Architecture-aware Robustness Evaluation of Explainable Deep Learning for Breast Cancer Diagnosis

Bridging visual saliency and large language models for explainable deep learning in medical imaging

Trusting Right Predictions for Wrong Reasons: A LIME Based Analysis of Deep Learning Interpretability in Lung Cancer Diagnosis