Feature-Sensitive Deep Convolutional Neural Network for Multi-Instance Breast Cancer Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Effective computer-aided detection (CAD) for breast cancer requires robust algorithms and well-labeled datasets.
• Mammography datasets often contain multiple views per case, with varying diagnostic relevance.

Purpose of the Study:

• To develop a feature-sensitive deep convolutional neural network (CNN) for improved breast cancer detection.
• To address the challenge of varying contributions from different mammographic views within a single case.

Main Methods:

• Construction of a multi-instance mammography clinic dataset with case-level pathological labels.
• Implementation of a feature-sensitive deep CNN using a pre-trained model for feature extraction.
• Development of a feature fusion module to dynamically weight instance features for classification.

Main Results:

• The proposed feature-sensitive deep CNN demonstrated effectiveness in breast cancer detection.
• Experimental results validated the method on a constructed clinic dataset and two public datasets.
• The feature fusion module successfully adapted instance contributions for improved classification.

Conclusions:

• The developed feature-sensitive deep CNN offers a promising approach for computer-aided breast cancer detection.
• Weighting features from different mammographic views enhances diagnostic performance.
• The end-to-end trained model shows significant effectiveness on diverse datasets.