Looking for Abnormalities in Mammograms With Self- and Weakly Supervised Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Millions of mammograms are generated annually, but lack detailed annotations, hindering abnormality detection.
• Current breast cancer screening protocols provide malignancy probability but not explicit region annotations, creating a data gap.

Purpose of the Study:

• To develop a robust abnormality detection framework for weakly annotated mammography datasets.
• To improve the reconstruction and classification of various breast abnormalities, including masses and micro-calcifications.

Main Methods:

• A mixed self- and weakly supervised learning framework was developed, combining domain knowledge with image-wise labels.
• High-resolution imaging was utilized to capture diverse findings like masses, micro-calcifications, distortions, and asymmetries.
• An auxiliary classification task was introduced for enhanced explainability based on reconstructed regions.

Main Results:

• The proposed method achieved an image-wise Area Under the Curve (AUC) of up to 0.86.
• An overall region detection true positive rate of 0.93 was recorded.
• A pixel-wise F1 score of 64% was obtained for malignant masses, demonstrating strong performance in segmentation and detection.

Conclusions:

• The developed framework effectively addresses the challenge of abnormality detection in weakly annotated mammograms.
• The method shows promise for improving the accuracy and explainability of AI in breast cancer screening.
• This approach advances the analysis of diverse findings in high-resolution mammographic images.