Enhancing Breast Lesion Detection in Mammograms via Transfer Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Early breast cancer detection via mammography significantly improves patient survival rates.
• Advanced object detection models are crucial for accurate analysis of mammographic images.

Purpose of the Study:

• To evaluate the performance of various object detection models (Cascade R-CNN, YOLOv12 variants, RTMDet-X, RT-DETR-X) for detecting breast cancer masses and calcifications.
• To assess the impact of standardized preprocessing and augmentation techniques on model performance.
• To investigate the effectiveness of transfer learning across different public mammography datasets.

Main Methods:

• Standardized preprocessing (CLAHE, cropping) and augmentation (rotations, scaling) were applied.
• Models were trained and validated on four public datasets: INbreast, CBIS-DDSM, VinDr-Mammo, and EMBED.
• Performance metrics included precision, recall, mean Average Precision at IoU 0.5 (mAP50), and F1-score.

Main Results:

• YOLOv12-L demonstrated excellent mass detection performance (mAP50=0.963, F1=0.917 on INbreast).
• Preprocessing improved mAP50 by up to 0.209; transfer learning boosted INbreast performance to mAP50=0.995.
• Domain shifts caused performance drops on CBIS-DDSM and VinDr-Mammo with transfer learning; calcification detection remained weak (mAP50 < 0.116).

Conclusions:

• High-capacity models, preprocessing, and augmentation are valuable for breast cancer mass detection.
• Further research is needed for robust calcification detection and effective domain adaptation in mammography AI models.