Mass segmentation for whole mammograms via attentive multi-task learning framework

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Accurate mass segmentation in mammograms is crucial for breast cancer diagnosis but challenging.
• Current methods often rely on manual or automatic patch extraction, which can be inefficient or error-prone.
• There is a need for improved, automated methods for precise mass segmentation in mammography.

Purpose of the Study:

• To develop a novel, end-to-end deep learning model for direct mass segmentation in whole mammograms.
• To improve the efficiency and accuracy of breast cancer mass segmentation, classification, and localization.
• To reduce errors associated with traditional patch extraction techniques.

Main Methods:

• Proposed an attentive multi-task learning network (MTLNet) for end-to-end mass segmentation.
• Utilized group convolution for efficient feature extraction and enhanced learning capacity.
• Incorporated an attention mechanism to focus on informative feature channels.
• Employed a multi-task learning framework for segmentation, classification, and localization to mitigate overfitting.

Main Results:

• The MTLNet achieved high performance on two public mammographic datasets (INbreast and CBIS-DDSM).
• Achieved a Dice index of 0.826 on the INbreast dataset for mass segmentation.
• Achieved a Dice index of 0.863 on the CBIS-DDSM dataset for mass segmentation.
• The multi-task approach enabled simultaneous segmentation, classification, and localization of masses.

Conclusions:

• The proposed attentive multi-task learning network (MTLNet) offers an effective and efficient solution for mass segmentation in mammography.
• MTLNet directly processes whole mammograms, eliminating the need for pre-processing patch extraction and its associated errors.
• The model's ability to perform segmentation, classification, and localization simultaneously enhances its utility in computer-aided diagnosis of breast cancer.