Breast Tumor Detection and Diagnosis Using an Improved Faster R-CNN in DCE-MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• AI-based breast cancer detection offers improved sensitivity and specificity, crucial for early diagnosis and mortality reduction.
• Existing two-stage detection networks struggle with imprecise Region of Interest (ROI) selection, leading to inclusion of surrounding tissues and false positives from fuzzy noise.

Purpose of the Study:

• To enhance AI-based breast cancer detection accuracy by addressing limitations in current two-stage networks.
• To minimize false positives and improve the precision of lesion identification in mammography.

Main Methods:

• Utilized Faster RCNN architecture with ROI aligning to reduce quantization errors.
• Incorporated Feature Pyramid Network (FPN) for multi-resolution feature extraction.
• Developed a bounding box quadratic regression network and convolutional layers to refine feature maps and reduce surrounding tissue interference.

Main Results:

• The proposed model outperformed Faster R-CNN, Mask R-CNN, and YOLOv9 on an internal dataset of 485 cases, achieving superior mAP, sensitivity, and false positive rates.
• Demonstrated a 38.5% reduction in false positives compared to manual detection.
• Achieved best performance on a public dataset of 220 cases, showing improved sensitivity and specificity.

Conclusions:

• The developed AI model significantly enhances breast cancer detection accuracy and reduces false positives.
• The approach effectively assists clinicians in diagnosing breast cancer, potentially improving patient outcomes through earlier and more precise detection.