Analyzing explainability of YOLO-based breast cancer detection using heat map visualizations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of mortality in women, necessitating early detection through mammography.
• Artificial intelligence (AI), particularly You Only Look Once (YOLO) models, shows promise in medical image analysis.
• Current AI models often lack transparency, hindering trustworthiness in clinical applications.

Purpose of the Study:

• To evaluate the performance of YOLOv9, YOLOv10, and YOLO11 for breast cancer detection using INbreast and MIAS datasets.
• To integrate and assess seven explainable artificial intelligence (XAI) methods with the best-performing YOLO model.
• To enhance the transparency and trustworthiness of AI-driven breast cancer detection.

Main Methods:

• Trained YOLOv9, YOLOv10, and YOLO11 on INbreast and MIAS mammography datasets.
• Selected the top-performing YOLO model based on mean average precision (mAP), precision, and recall.
• Integrated seven XAI methods (Grad-CAM, Grad-CAM++, Eigen-CAM, EigenGrad-CAM, XGrad-CAM, LayerCAM, HiResCAM) for model interpretability.

Main Results:

• YOLO11 achieved the highest mAP (0.935), outperforming YOLOv9 (0.868) and YOLOv10 (0.926).
• YOLO11 demonstrated classification accuracies of 95% for benign and 80% for malignant cases.
• HiResCAM provided the most effective visual explanations, achieving the highest matching ground truth (mGT) score of 0.49.

Conclusions:

• The combination of YOLO11 and HiResCAM offers high accuracy and improved interpretability for breast cancer detection.
• This integrated approach enhances user trust by visualizing AI decision-making processes.
• The findings provide insights for refining AI models and improving their performance in medical diagnostics.