Explainable machine learning for breast cancer diagnosis from mammography and ultrasound images: a systematic review

Introduction

Breast cancer is the first and most common type of cancer in women.1 2 Anatomically, the breast consists of healthy blood vessels, connective tissue, ductal lobules and lymph nodes.3 Breast cancer is a problem with abnormal growth of the breast cells. By 2040, the burden of breast cancer is predicted to increase to over three million new cases and one million deaths every year because of population growth and ageing alone.2

Breast cancer is highly treatable if identified at an early stage, and hence, early detection is crucial to save lives. Among the methods of breast cancer detection, the most popular are ultrasound (US),4 mammography5 and MRI. However, traditional computer-aided design systems generally depend on manually created features and experience of the physiologist, therefore weakening the overall performance of breast cancer identification. Therefore, artificial intelligence (AI) methods like machine learning and deep learning-based techniques have emerged for breast cancer diagnosis with high accuracy. Additionally, improved breast cancer classification by combining graph convolutional network and convolutional neural network6 and abnormal breast identification by a nine-layer convolutional neural network with parametric rectified linear unit and rank-based stochastic pooling are used to support patients and doctors’ decisions.7 However, the algorithms lack ethical AI, right of explanation and trustworthy AI. These concepts are considered critical issues by high-level political and technical bodies (eg, G20, EU expert groups, Association of Computing Machinery in the USA).8 9

Additionally, AI algorithms like machine learning and deep learning are vulnerable to bad stuff (bad decisions, bad medical diagnosis and bad prediction) is the most common drawback of AI algorithms today. They are also black box for predictive interpretation.

To overcome this issue, the science of explainable AI (XAI) has grown exponentially with its successful application in breast cancer diagnosis. However, it still requires a comprehensive review of existing studies to help researchers and practitioners gain insight and understanding of the field. Therefore, his systematic review is conducted.

XAI is the extent to which people can easily understand the model. It has received much attention over the past few years. The purpose of a model explanation is to clarify why the model makes a certain prediction, to increase confidence in the model’s predictions10 and to describe exactly how a machine learning model achieves its properties.11 Therefore, using machine learning explanations can increase the transparency, interpretability, fairness, robustness, privacy, trust and reliability of machine learning models. Recently, various methods have been proposed and used to improve the interpretation of machine learning models.

There are different taxonomies for machine learning explainability. An interactive explanation allows consumers to drill down or ask for different types of explanations until they are satisfied, while a static explanation refers to one that does not change in response to feedback from the consumer.12 A local explanation is for a single prediction, whereas a global explanation describes the behaviour of the entire model. A directly interpretable model is one that by its intrinsic transparent nature is understandable by most consumers, whereas a post hoc explanation involves an auxiliary method to explain a model after it has been trained.13 Self-explaining may not necessarily be a directly interpretable model. By itself, it generates local explanations. A surrogate model is usually a directly interpretable model that approximates a more complex model, while visualisation of a model may focus on parts of it and is not itself a full-fledged model.

No single method is always the best for interpreting machine learning.12 For this reason, it is necessary to have the skills and equipment to fill the gap from research to practice. To do so, XAI toolkits like AIX360,12 Alibi,14 Skater,15 H2O,16 17 InterpretML,18 19 EthicalML-XAI,19 20 DALEX,21 22 tf-explain,23 Investigate.24 Most interpretations and explanations are post hoc (local interpretable model-agnostic explanations (LIME) and SHapley Additive exPlanations (SHAP). LIME and SHAP are broadly used explanation types for machine learning models from physical examination datasets. But these made explanations with limited meaning as they lacked fidelity and transparency. However, deep learning and ensemble gradients are preferable in performance for image processing and computer vision. This research is processing mammography and US images. Therefore, deep learning is recommended for breast cancer image processing.

Ensemble gradients are used to interpret deep neural networks,11 GradientSHAP is a sample interpretation algorithm that approximates SHAP values.25 Occlusion methods are most useful in situations such as image processing. Biological nurturing(BN) is ideal for clinical decision-making and, in general, for all assessments and studies involving multiple interventions and orientations. The oriented, modified integrated gradient (OMIG) interpretability method is inspired by the integrated gradients method. Since there is no one-size-fits-all approach to learning machine explanation, it needs a comprehensive evaluation of published papers and tools to bridge the gap in research to practice.

The research that does not consider objective metrics for evaluating XAI may lack significance and experience controversy, especially if negative reviews are not used.8 To avoid the issues, a study8 suggests four metrics based on performance differences, D, between the explanation’s logic and the agent’s actual performance, the number of rules,    outputted by the explanation, the number of features, F, used to generate the explanation, and the stability, S, of the explanation. It is believed that user studies that focus on D, R, F and S metrics in their evaluations are inherently more valid.

The main contributions of this systematic review are:

1. Investigating XAI methods popularly applied for breast cancer diagnosis.

2. Identifying the algorithm’s explainability and their performance relation in breast cancer diagnosis.

3. Summarise the evaluation metrics used for breast cancer diagnosis using XAI methods.

4. Summarise existing ethical challenges that XAI overcomes in breast cancer diagnoses.

5. Analysing the research gaps and future direction for XAI for breast cancer detection.

Methodology

The methodology employed in this systematic review is devoid of any medical (either prospective or retrospective) data of patients. This study applies the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guiding principles for conducting systematic reviews.26 PRISMA 2020 was adopted because of the clear guidelines it offers to ease robust systematic reviews. Therefore, this review article follows the recommendations of the guidelines. There is no stated date limit to filter the papers. The papers were searched on 19 September 2023. Peer-reviewed manuscripts and conference proceedings from PubMed, IEEE Explore, ScienceDirect, Scopus and Google Scholar databases published were searched. Rayyan for systematic review was used for duplicate removing, inclusion and exclusion term visualisations. The systematic review protocol was registered through PROSPERO with ID CRD42023458665.27 Preplanned subgroup analyses were detailed.

Study selection

The selection process of the articles was conducted based on the inclusion and exclusion criteria defined (figure 1). A bibliography of 646 papers was extracted from databases (PubMed=118, ScienceDirect=331, Scopus=88, Google Scholar=102 and IEEE Xplore=7). All the extracted papers were imported into the Rayyan online platform for systematic review. In total, 132 articles were found to be duplicates and were deleted. Moreover, 501 papers were excluded (systematic review, scoping review, breast cancer diagnosis without explainable AI and explainable AI without breast cancer diagnosis). In total, 79 papers with XAI for breast cancer terms were retained. Their full documents were downloaded and reviewed. From these, 65 papers with XAI for breast cancer without mammography or US terms were excluded again. Finally, 14 studies with XAI for breast cancer and mammography or US or both terms were included and used for this systematic review.

Request permission

Figure 1

Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow chart of explainable artificial intelligence (XAI) for breast cancer diagnosis.

Results and discussion

Results

A total of 646 papers were extracted using search queries and terms defined in tables 1 and 2 from the selected databases. From a total of 646 papers, 134 were duplicates and removed. As depicted in figure 1, based on inclusion and exclusion stated in section Inclusion and exclusion criteria above, 79 papers (14%) with XAI for breast cancer were included (figure 1). Figure 2 depicts the included and excluded ratios. All screenshots added to these results are taken from Rayyan for a systematic review online platform.

Request permission

Figure 2

Included and excluded ratio graph for explainable artificial intelligence, breast cancer and mammography or ultrasound.

US and mammography are the most recommended methods for breast cancer diagnosis. From 79 included papers based on XAI for breast cancer, 14 papers with XAI for breast cancer and mammography or US or both terms were either included or excluded based on criteria set in section Inclusion and exclusion criteria above. So, table 3 presents that 64.29% (9 papers from included 14) of papers were on US images, whereas 35.71% (5 papers from included 14) of papers were on mammography images.

Table 3

View inline|View popup

Overview of reviewed articles on explainable artificial intelligence data

Figure 2 shows that 97% were excluded and 3% were included based on inclusion criteria. Table 3 shows that 100% of the included papers visualised are XAI for breast cancer from mammography, US or both. It shows that 50% of them used heatmaps for visualisation.

The main objective of XAI is to encounter ethical challenges and to increase doctors’ and patients’ thrust on XAI. Different XAI are used for breast cancer. However, only one paper compared doctors’ trust in the system.

In most of the papers, 50% (7 from 14 papers) used heatmaps for visualisation of areas of interest29–35 and.36 Additionally, Zhang et al37 used BI-RADS-Net, Zhang et al38 and Shen et al35 used a saliency map, Ortega-Martorell et al39 used uniform manifold approximation and projection (UMAP), Mital and Nguyen40 used a tornado diagram, Rezazadeh et al41 used histogram and Rezazade Mehrizii et al34 used class activation map (CAM)-based heatmaps.

Shen et al’s study35 used the largest number of datasets when compared with included studies. The study proves that the artificial intelligence system reduces false-positive findings in the interpretation of breast US examinations.35 Breast cancer is most common in women, based on evidence on the ground in all of the studies most of the data are from women. This implies the ground truth. However, most of the datasets are taken from women and do not keep the existence of breast cancers in the ratio from man to women.

A total of 5 648 066 datasets are used by all included papers. From all the included papers, US-based datasets were used by 99% of studies. Mammography-based datasets used by only 1% of the total studies. For example, the maximum datasets used by Shen et al35 used 5 442 907 US images, and the study by Mital and Nguyen40 used 100 000 mammography images. This shows that there are many works left to work on improving the number of datasets on mammography images when compared with US images. We recommend that data should be collected from suspected patients with breast cancer but all the included studies said nothing about it.

Explainable/interpretable algorithms used are deep learning explanation algorithms: Of 14 papers, Explainer alone or with Grad-CAM,29 interpretable deep learning,30 Grad-CAM,31 Fisher information network (FIN),39 AI and Polygenic Risk Scores (PRS) algorithms,40 DenseNet,35 Explainability-partial,34 Explainability-full,34 VGG-16,37 fine-tuned MobileNet-V2 convolutional neural network,33 OMIG explainability32 and BI-RADS-Net-V238 are used in 11 papers (78.57 %), SHAP41 ,42 is used in 2 papers (14.3%) and LIME36 is used in 1 paper (7.14%).

Risk of bias

The study population was known in all articles. We have obtained complete outcome variables in all articles. In all articles involved, selective reporting and publication bias were not obtained (figure 3). ‘Traffic light’ plots of the domain-level judgments for each result are sh0wn in figure 3.

Request permission

Figure 3

Traffic light plot for risk of bias.

Discussion

Explainer is the situation that is explainable by itself rather than explaining black box.29 They proved that physicians perform better when assisted by Explainer than when diagnosing alone. The study compares the use of Explainer with the post hoc technique. Based on this, they prove that Explainer can locate more reasonable and feature-related regions than the classic post hoc technique. Robustness is a characteristic expected from XAI. The study by Song et al29 also tested the robustness of the proposed framework. Explainability29 is not only related to AI performance but also to responsibility and risk in medical diagnosis. For phantom object detection,30 accuracy and mean intersection over union were used to test the model over a total of 6369 out of 6400 objects. Finally, Oh et al’s study30 concludes interpretable deep learning model using large-scale data from multiple centres shows high performance.

In the study by Qian et al,31 BI-RADS scores for breast cancer were compared with experienced radiologists, areas under the receiver operating curve (ROC) and CI for multimodal images. Explanation using principal component analysis, visualisation using UMAP, FIN visualisations of the training cases and projecting the test cases onto the trained embedding.39 the study propose a novel visualisation using FIN containing accurate information about data points’ similarities that can provide intelligence about neighbouring data points.

The finding by Mital and Nguyen40 explained AI’s ability to identify high-risk women more accurately than PRS, and family history reduces the possibility of delayed breast cancer diagnosis and fewer false-positive diagnoses from not screening low-risk women.

In Sun et al’s study,42 model-agnostic methods versus model-specific methods, post hoc (black box+SHAP) technique and three algorithms, namely, logistic regression, extreme gradient boosting and random forest performance, were evaluated by sensitivity, specificity and AUC.42 This evaluation was used to evaluate the black box model only. Moreover, SHAP was used for visualising feature importance using a heatmap but it was not tested.

In Lee et al’s study,36 accuracy, sensitivity, specificity and AUC were used. Simple linear iterative clustering superpixel segmentation method and the LIME explanation algorithm were employed to explain how the model makes decisions.

The area under the ROC of machine learning and an average of 10 board-certified breast radiologists were compared.35 In this case, radiologists decreased their false-positive rates with the help of XAI. They also evaluated an independent external test dataset to prove the potential of XAI in improving the accuracy, consistency and efficiency of breast US diagnosis worldwide. The study 35 discuss accuracy of the VGG backbone to ResNet50 and EfficientNet B0 backbone was evaluated and BI-RADS descriptors were used to evaluate.37

In Zhang et al’s study,38 accuracy, sensitivity, specificity, F1 score, R2, Mean Squared Error (MSE), Root Mean Square Error (RMSE),d shape orientation and margin were used to test the likelihood of malignancy. Explainer I was used to explain the classification results semantically. Explainer II constructs a quantitative explanation based on the classifier and Explainer I.

The study by Amanova et al32 proposes and applies a new explainability method: OMIG method. The study proved that the proposed approach yields substantially more expressive and informative results for our specific use case. To avoid issues like limited meaning and confirmation bias due to low-fidelity explanations unnecessarily, Gurmessa and Jimma8 suggest four metrics based on performance (D, R, F and S), but none of the selected studies used these metrics.

Bad stuff (bad decision, bad medical diagnosis and bad prediction) is the most common drawback of AI algorithms today. However, XAI could resolve this drawback. Robustness is also a characteristic expected from XAI. The study by Song et al29 tested the robustness of the proposed framework. This study puts explainability as not only related to AI performance but also to responsibility and risk in medical diagnosis. XAI proves that the performance of algorithms is complementary but not enough alone. The complementing of both performance and explainability satisfaction increases the system’s acceptance of legal and personal recognition.

Conclusion

In addition to increasing accuracy, reducing human error and technological advancement, XAI for breast cancer diagnosis overcomes ethical challenges by providing the right to know, robustness, transparency, accountability and interpretability in the decision-making process of machine learning models. However, it is not approved that it increases users’ and doctors’ trust in the system. Effective and systematic evaluation of its usefulness in this scenario is also lacking. Additionally, further work is needed to enhance the interpretability of deep learning algorithms through overcoming explainable to accuracy trade-offs, as well as to investigate the potential insights they can provide for clinicians’ decision-making.