Machine Learning-Driven Breast Cancer Diagnosis Software Integrated with Explainable Artificial Intelligence Based on Fine Needle Aspirate Findings

Breast cancer is typically the most frequently diagnosed malignancy and the prime reason for cancer-related deaths in women worldwide hence a precise determination of breast cancer is crucial. Numerous research studies have been published regarding automated diagnosis of breast cancer with the help of machine learning models. However, most of these studies suffer from lack of transparency behind the reasoning used by the models in coming up with the prediction resulting to lack of trust among physicians and patients. Thus, the aim of the study is to create a decision support tool that can predict the existence of breast cancer using machine learning classification models while incorporating an explainable artificial intelligence component. Random forest, logistic regression, support vector machine, and k-nearest neighbors were applied and evaluated on an anonymized freely accessible breast cancer dataset. Additionally, feature scaling to normalize the values, random oversampling to handle class imbalance, and feature extraction using principal component analysis to identify the groups of features that are related to each other were done. Overall, the best model was random forest with feature scaling and random oversampling having an area under the ROC curve = 0.9991, acc = 0.9832, precision = 0.9731, and F1-score = 0.9821. The best model together with an explainable artificial intelligence component using Local Interpretable Model-agnostic Explanations was integrated into a simple prototype application, “BreastFriend”, which enabled a pathologist to forecast the existence of breast cancer given a set of fine needle aspiration cytology features. It provided an explanation why the model made a particular prediction thus can serve as a decision support tool for physicians on the appropriate therapeutic options for the patient. Prompt and accurate diagnosis of breast cancer is within reach through an active collaboration between healthcare practitioners and machine learning experts leading to a more personalized treatment options and thus improve survivorship of the patients.