Enhanced Multiple Instance Learning for Breast Cancer Detection in Mammography: Adaptive Patching, Advanced Pooling,...

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Weakly supervised learning presents challenges in medical image analysis.
• Accurate breast cancer detection in mammography requires robust feature learning.
• Existing methods often need detailed region-of-interest annotations.

Purpose of the Study:

• To develop an Enhanced Embedded Space Multiple Instance Learning Network (MI-Net) for weakly supervised breast cancer detection.
• To improve feature learning and classification performance using deep supervision.
• To evaluate the efficacy of various pooling methods within the MI-Net framework.

Main Methods:

• Implemented an Enhanced Embedded Space MI-Net incorporating adaptive patch creation and convolution feature extraction.
• Integrated multiple pooling strategies: max, mean, log-sum-expo, attention, and gated attention pooling.
• Employed deep supervision to enhance feature representation across network layers.

Main Results:

• The Enhanced MI-Net model with attention pooling achieved the highest Area Under the Curve (AUC) of 86% on the CBIS-DDSM dataset.
• Deep supervision significantly improved bag-level classification performance.
• Attention pooling demonstrated superior performance compared to other pooling methods.

Conclusions:

• The Enhanced Embedded Space MI-Net offers a robust and scalable solution for breast cancer detection in mammography.
• The model effectively leverages weakly supervised learning, reducing the need for detailed annotations.
• This approach shows significant clinical relevance as an efficient diagnostic tool for mammographic image analysis.