Weakly Supervised Breast Lesion Detection in Dynamic Contrast-Enhanced MRI

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Accurate medical annotations for tumor detection are labor-intensive, hindering supervised learning development.
• Dynamic Contrast-Enhanced MRI (DCE-MRI) is crucial for breast lesion visualization.
• Weakly supervised learning offers a potential solution to annotation limitations.

Purpose of the Study:

• To develop and evaluate a weakly supervised learning model for breast lesion detection using only image-level labels in DCE-MRI.
• To assess the model's performance in terms of classification and localization accuracy.
• To investigate the utility of Feature Pyramid Network (FPN) and Layer-CAM in this context.

Main Methods:

• A dataset of 652 DCE-MRI cases (254 normal, 398 abnormal) was utilized.
• A Feature Pyramid Network (FPN) with Convolutional Block Attention Module (CBAM) integrated with VGG16 was employed for feature extraction.
• Layer-CAM generated initial localization heatmaps, refined by Conditional Random Field (CRF).

Main Results:

• The model demonstrated high image-level classification performance: 95.2% accuracy, 91.6% sensitivity, 99.2% specificity, and 0.986 AUC.
• Breast lesion localization achieved an Average Precision (AP) of 84.1% using weakly supervised learning.
• The combined approach of FPN and Layer-CAM facilitated automatic breast lesion detection.

Conclusions:

• Weakly supervised learning, utilizing FPN and Layer-CAM, effectively detects breast lesions in DCE-MRI with image-level labels.
• This approach significantly reduces the reliance on extensive manual annotations, accelerating AI development in breast cancer diagnostics.
• The model's high accuracy and localization performance show promise for clinical application in breast lesion detection.