Detection and Classification of Lesions in Mammograms using One-Stage Models

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Breast cancer is the most common cancer in women, making early detection crucial.
• Computer-Aided Diagnosis (CAD) systems aim to improve lesion detection in mammograms.
• Artificial Intelligence (AI) in radiology shows potential for enhancing diagnostic accuracy.

Purpose of the Study:

• To compare object detection models for smart diagnostic systems in mammography.
• To evaluate the You Only Look Once version 12 (YOLO-v12) architecture for automated lesion detection, localization, and malignancy assessment.
• To benchmark YOLO-v12 against Detection Transformer (DETR) and RetinaNet for mammographic analysis.

Main Methods:

• A comparative experimental study using retrospective data.
• Training and testing models on the Categorized Digital Database for Low-Energy and Subtracted Contrast-Enhanced Spectral Mammography (CDD-CESM) dataset.
• Utilizing 1,982 mammograms with 3,720 annotated lesions for model evaluation.

Main Results:

• YOLO-v12 achieved excellent diagnostic accuracy with a mean Average Precision (mAP50) of 0.98 and Intersection Over Union (IOU) of 0.95.
• YOLO-v12 significantly outperformed contemporary models and previous YOLO versions.
• The model demonstrated high precision in detecting and localizing lesions and assessing their status.

Conclusions:

• AI technologies, particularly YOLO-v12, show significant potential to assist radiologists in early breast cancer detection.
• The findings support the implementation of YOLO-v12 in clinical mammography screening.
• Future research should focus on real-time diagnostic systems to further improve breast cancer detection capabilities.