Explainable artificial intelligence in breast cancer detection and risk prediction: A systematic scoping review

1 | INTRODUCTION

• Early detection and treatment increase survival chances, while AI-based computer-aided diagnosis supports cancer detection, classification, and diagnosis of biomarkers and morphological features.Source: §1 · p. 2
• AI can learn patterns from medical data, but bias and opacity raise accuracy, accountability, fairness, and transparency concerns in high-stakes healthcare.Bias may stem from low-quality datasets, faulty algorithms, or human cognitive biases.Source: §1 · p. 2
• XAI aims to make black-box algorithms more understandable and interpretable, on the assumption that transparency can support trustworthy, safe, and reliable decisions.Source: §1 · p. 2
• In the human-in-the-loop process, clinicians and XAI experts use explanations to validate, confirm, or enhance AI predictions, with iterative review when results are unsatisfactory.The described workflow can revisit the model, data, or both before returning outcomes to the AI system.Source: §1 · p. 2
• Where earlier reviews largely covered general healthcare, this review summarizes XAI studies in breast-cancer screening, risk detection, and prediction and discusses methods and their popularity.Source: §1 · pp. 2–3

2.1 | Overview of AI models

• Machine learning covers supervised regression and classification and unsupervised clustering, while deep learning may be supervised or unsupervised.Regression predicts continuous outputs, classification predicts discrete labels or categories, and clustering groups similar data points.Source: §2.1 · p. 3
• Deep learning automates feature extraction and can model complex nonlinear relationships in breast-cancer data that are high-dimensional, multimodal, noisy, and sparse.The review describes deep learning as suitable for unstructured data such as images, whereas classical machine learning does not function effectively on such data.Source: §2.1 · p. 3
• Ensemble learning combines multiple models to make a decision, typically in supervised machine-learning tasks.Source: §2.1 · p. 3

2.2 | Accuracy‐explainability trade‐off

• The accuracy–explainability trade-off is the balance between a model’s predictive accuracy and explainability, and achieving both is not straightforward.Source: §2.2 · p. 4
• White-box models are intrinsically transparent but limited to linear associations, black-box models are nontransparent, and gray-box models balance accuracy and explainability.Gray-box connections from inputs to outputs can be explained even when the model is not fully transparent.Source: §2.2 · p. 4

2.3 | Classification of XAI methods

• XAI methods can be classified by scope (local or global), explanation stage (intrinsic or post hoc), and explanation type.Explanation types include feature importance, white-box models, example-based methods, and visual explanations.Source: §2.3 · pp. 4–5

• Following PRISMA guidance, two reviewers extracted eight study characteristics, including authorship, year, aims, datasets, data types, important features, AI type, and the explained model.Disagreements were resolved by consensus.Source: §3 · pp. 5–6
• The review searched Scopus, IEEE Xplore, PubMed, and the first 50 Google Scholar citations from January 2017 to July 2023 for studies of XAI in breast cancer.The search used combinations of keywords and MeSH terms.Source: §3 · pp. 5–6
• Of 193 studies identified, 30 met the inclusion criteria after screening for original, peer-reviewed English-language journal studies using at least one XAI method in breast cancer.After title-and-abstract screening, 49 articles underwent full-text scrutiny; exclusions included reviews, conference papers, duplicates, and unrelated or nonresearch material.Source: §3 · p. 5

• Almost 30 studies used XAI methods in breast cancer settings.Source: §4 · p. 7
• The identified methods included SHAP, LIME, CAM, Grad-CAM, Grad-CAM++, and LRP.Source: §4 · p. 7
• Figure 5 shows study and model counts by method, while Tables 4–9 tabulate the included studies.Source: §4 · p. 7

5.1.1 | SHapley Additive exPlanations (SHAP)

• SHAP treats observation features as game players and predictions as rewards, assigning average marginal contributions for local and global explanations.The review describes this as a fair distribution of reward among players.Source: §5.1.1 · p. 7
• Twelve of 13 SHAP studies used ensemble machine learning; one study used logistic regression with IHC and FISH diagnoses as training outputs.Source: §5.1.1 · p. 7
• SHAP linked B-cell and CD8+ T-cell or NK T-cell fractions above inflection points with up to 18% higher survival, and highlighted GLCM texture features.The ultrasound analysis identified GLCM correlation and energy at different pixel distances along the 90° direction.Source: §5.1.1 · p. 7
• SHAP was used in 13/30 studies; no deep-learning models were paired with it, and XGBoost appeared in 9/13 studies.The review attributes SHAP’s fit with tree-based models to its high-speed algorithm [18].Source: §5.1.1 · p. 7

5.1.2 | Local interpretable model agnostic explanations (LIME)

• LIME appeared in 5/30 studies and built local linear surrogate explanations by perturbing features and observing output changes.For images, perturbations can replace image regions with gray; one example used yellow pixel masks to highlight image segments.Source: §5.1.2 · p. 7
• With SHAP and LIME, compact biological gene clusters achieved similar or better performance than larger gene sets, and LIME improved AUC except for RF and rSVM.Source: §5.1.2 · p. 7
• LIME explained distant-metastasis predictions with impacts ranging from high for nonuse of adjuvant chemotherapy to low for oral contraception.In another comparison, RF focused on nine features, while the nonlinear comparator used all but one feature.Source: §5.1.2 · p. 12
• Across 18/30 studies using model-agnostic methods, SHAP was preferred for feature importance; LIME remained a local option and can be faster for many predictions.The review cites SHAP’s local and global explanations and faster global explanations for high-performance ensembles as reasons for its preference.Source: §5.1.2 · p. 12

5.2 | Post hoc XAI methods: Model‐ specific

• Most deep-learning-specific explainers are propagation-based and use gradients available during training.Source: §5.2 · p. 12

5.2.1 | Class activation map (CAM)

• CAM uses a GAP layer and classification layer after the final convolution to localize discriminative image regions, but requires that architecture.Five of 30 studies used CAM to assess CNN breast-mass localization.Source: §5.2.1 · p. 12

5.2.2 | Gradient‐weighted class activation mapping (Grad‐CAM)

• Grad-CAM creates a coarse, class-specific heatmap from a CNN’s last-layer feature maps and works with differentiable CNN architectures without retraining or modification.Four of 30 studies used Grad-CAM to assess CNN breast-mass localization.Source: §5.2.2 · p. 13
• Grad-CAM++ extends Grad-CAM for improved object localization and multiple same-class objects; one study used it on DUV whole-slide images with pretrained DenseNet169.The study generated regional significance maps to classify each slide as cancerous or benign.Source: §5.2.3 · p. 13

5.2.4 | Layer‐wise relevance propagation (LRP)

• LRP propagates relevance scores backward from a classifier output through deep-learning layers to the input, producing a heatmap of critical image regions.Two studies implemented LRP.Source: §5.2.4 · p. 14
• LRP-based relevance aggregation scores input features across samples, while GLRP produces patient-specific molecular subnetworks that agree with clinical knowledge and identify potential tumor-progression drivers.Source: §5.2.4 · p. 14
• 12 of 30 studies used model-specific XAI methods, with CAM and Grad-CAM the most used.Source: §5.2.4 · p. 14

• Studies focused mainly on breast-cancer diagnosis/classification or patient survival/prognosis, with SHAP more often used for clinical data than image recognition.Figure 6 shows the distribution of XAI methods across clinical applications, with bubble numbers representing study counts.Source: §5.3 · p. 14, Figure 6 · p. 14
• CAM and Grad-CAM are less computationally intensive than SHAP, making them more suitable for image processing, especially in real-time applications.Source: §5.3 · p. 14
• Diagnostic studies used supervised learning to distinguish healthy from diseased patients, and XAI helped clinicians understand and validate features influencing outcomes.Source: §5.3 · p. 14
• Survival and prognosis models predicted events such as mortality or metastasis, while XAI interpreted how factors contributed to patient outcomes.The review says this interpretability can make models more understandable, usable, and trustworthy to clinicians and patients.Source: §5.3 · p. 14

5.4 | Future directions

• Future research should refine XAI methods to explain increasingly intricate healthcare AI models, including advanced generative models.Source: §5.4 · p. 14

5.5 | Study limitations

• Excluding non-English, inaccessible, and gray-literature articles may have omitted relevant information.Source: §5.5 · p. 15
• The review focused on established XAI methods and excluded studies with few or no citations; terminology diversity may also have caused studies to be missed.Source: §5.5 · p. 15

• SHAP was the most-used model-agnostic method, often paired with tree-based ensemble models for their speed and compatibility; CAM and Grad-CAM were widely used model-specific methods.Source: §6 · p. 15
• The reviewed XAI methods provided only sanity checks, and oncologists did not evaluate their outputs; clinical evaluation is needed to assess reliability.Source: §6 · p. 15
• Multimodal XAI faces challenges including data resolution and representativeness, heterogeneous modalities, dataset readiness, and algorithmic or human biases in explanations.The review identifies addressing these challenges as key to widespread acceptance in cancer care delivery and treatment.Source: §6 · p. 15

CONFLICT OF INTEREST STATEMENT

• The authors declare no conflict of interest.Source: CONFLICT OF INTEREST STATEMENT · p. 15

DATA AVAILABILITY STATEMENT

• The authors state that the data supporting the study’s findings are available within the paper.Source: DATA AVAILABILITY STATEMENT · p. 15

ORCID

• ORCID identifiers are provided for Amirehsan Ghasemi, Soheil Hashtarkhani, David L. Schwartz, and Arash Shaban‐Nejad.Source: ORCID · p. 16

APPENDIX A: POPULAR MODEL‐ AGNOSTIC AND MODEL‐SPECIFIC XAI METHODS

• Appendix A presents separate tables for model-agnostic and model-specific XAI methods.Source: Table A1 · p. 21, Appendix A · p. 21, Table A3 · p. 22
• Separate appendix tables list code or toolboxes for the model-agnostic and model-specific methods.Source: Table A2 · p. 21, Appendix A · pp. 21–22, Table A4 · p. 22
• The listed code/toolbox entries include PDP and guided backpropagation.Source: Appendix A · pp. 21–22