Enhanced multiple instance learning for breast cancer detection in mammography: Adaptive patching, advanced pooling, and deep supervision

Abstract

This paper addresses the challenge of weakly supervised learning for breast cancer detection in mammography by introducing an Enhanced Embedded Space MI-Net model with deep supervision. The framework integrated adaptive patch creation, convolution feature extraction, and pooling methods -max, mean, log-sum-expo, attention, and gated attention pooling - evaluated in three MIL models, Instance Space mi-Net, Embedded Space MI-Net and Enhanced Embedded Space MI-Net. A key contribution is the incorporation of deep supervision, improving feature learning across network layers and enhancing bag-level classification performance. Experimental results on the CBIS / DDSM dataset demonstrate that the Enhanced MI-Net model achieves the highest AUC of 86% with attention pooling. This work addresses the gap in leveraging MIL techniques for high-resolution medical imaging without requiring detailed annotations, offering a robust and scalable solution for breast cancer detection.Clinical Relevance - This study highlights the potential of MIL-based models with attention pooling to accurately detect breast cancer in mammographic images without requiring detailed ROI annotations, offering a scalable and efficient diagnostic tool for clinical practice.