This paper presents DXAIB - a hybrid system that combines a Convolutional Neural Network (CNN) with a Random Forest classifier to detect breast cancer from medical data. The goal is not just to make accurate predictions, but also to explain why the system makes each decision.
A key challenge in medical AI is the 'black box' problem - powerful models make predictions without showing their reasoning. Doctors need to understand and trust AI recommendations before acting on them. DXAIB addresses this by adding SHAP (SHapley Additive exPlanations), a tool that shows which features most influenced each prediction.
The system was tested on the widely used Wisconsin Breast Cancer Dataset, which contains measurements from fine needle aspiration biopsies of breast masses. This is a well-established benchmark for comparing classification methods.
Breast cancer is one of the most common and deadly cancers affecting women worldwide. Early and accurate diagnosis dramatically improves survival rates, making reliable automated tools valuable for supporting clinicians.
Traditional deep learning models - especially neural networks - can achieve high accuracy but are difficult to interpret. Radiologists and oncologists are unlikely to adopt tools they cannot understand or verify. Explainable AI (XAI) bridges this gap by surfacing the internal logic of complex models.
SHAP values borrow from game theory: they treat each feature as a 'player' and fairly distribute the prediction outcome among all features based on their contribution. This gives a clear, consistent picture of what drove each individual diagnosis.
The system works in two stages. First, a CNN extracts deep feature representations from the input data - essentially learning which combinations of measurements are meaningful for distinguishing malignant from benign tumors. CNNs are excellent at finding patterns that human-engineered features might miss.
Second, the extracted CNN features are passed to a Random Forest classifier, which makes the final malignant/benign prediction. Random Forests combine many decision trees and are naturally more interpretable than pure neural networks while maintaining strong predictive performance.
After prediction, SHAP is applied to explain each result. The SHAP analysis shows which input features (such as cell radius, texture, or concavity) pushed the prediction toward malignant or benign, and by how much. This explanation is generated for every individual patient prediction.
DXAIB achieved 98.35% accuracy on the Wisconsin Breast Cancer Dataset - meaning it correctly classified nearly all cases. The system also achieved 98.76% precision (very few false positives), 98.74% recall (very few missed cancers), and a 98.72% F1 score, which balances precision and recall.
These results outperformed or matched competing methods tested on the same dataset, including standalone CNNs, SVMs, and other ensemble approaches. The hybrid CNN + Random Forest combination proved superior to using either component alone.
Importantly, high recall (sensitivity) is especially critical in cancer screening - missing a true cancer is more dangerous than a false alarm. The near-99% recall means almost no cancers were incorrectly classified as benign.
The SHAP analysis identified which cell measurements most consistently separated malignant from benign tumors. Features related to cell size (radius, area, perimeter) and shape irregularity (concavity, compactness) were the strongest predictors of malignancy.
SHAP summary plots showed that larger, more irregular cells with higher concavity values strongly pushed predictions toward malignant, while smaller, more regular cells pointed toward benign. These patterns align with what pathologists look for in microscopy.
This agreement between the AI's learned features and established medical knowledge is a strong sign that DXAIB is capturing clinically meaningful signals rather than statistical artifacts.
The DXAIB system is designed as a decision-support tool - it assists radiologists and pathologists rather than replacing them. By providing both a prediction and an explanation, it gives clinicians the information they need to agree, question, or override the system's recommendation.
The explainability layer could be particularly valuable in borderline cases where the diagnosis is uncertain. A doctor can review which features the AI found unusual and focus their own examination on those areas.
Integrating such tools into clinical workflows could reduce diagnostic variability between different clinicians and help less-experienced practitioners make more consistent decisions.
DXAIB demonstrates that it is possible to achieve near-perfect classification accuracy while simultaneously providing clear, interpretable explanations. The two goals - performance and explainability - are not mutually exclusive.
The authors argue that future medical AI systems should routinely include explainability components. As AI becomes more embedded in diagnostic pathways, the ability to audit and understand model behavior becomes a patient safety issue, not just a technical nicety.
Future work could extend this approach to imaging data such as mammograms or histology slides, where CNN feature extraction is already standard practice but explainability remains an open challenge.