Local Extremum Mapping for Weak Supervision Learning on Mammogram Classification and Localization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Early and accurate breast lesion detection via mammography is vital for patient survival.
• Current deep learning methods require expensive pixel-level annotations, hindering real-world application.
• Weakly supervised learning offers a potential solution to reduce annotation costs.

Purpose of the Study:

• To propose a novel Local Extremum Mapping (LEM) mechanism for mammogram classification and weakly supervised lesion localization.
• To reduce the reliance on pixel-level annotations in deep learning models for mammography.
• To enhance the efficiency and accessibility of automated mammogram analysis.

Main Methods:

• Dividing mammograms into regions and generating score maps using convolutional neural networks.
• Identifying informative regions by filtering local extrema in score maps.
• Aggregating scores from informative regions for final classification and lesion localization.

Main Results:

• Achieved competitive performance on CBIS-DDSM and INbreast datasets.
• Improved classification accuracy to 96.3% with an AUC of 0.976 on the INbreast dataset.
• Effectively localized lesions with a Dice Similarity Coefficient of 0.37, outperforming baseline methods like Grad-CAM.

Conclusions:

• The proposed LEM method enables accurate mammogram classification and lesion localization using only image-level labels.
• This approach significantly reduces annotation costs, offering practical clinical significance.
• LEM enhances the accessibility and efficiency of automated mammogram analysis for potential clinical applications.