Convolutional Feature Descriptor Selection for Mammogram Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading global cancer diagnosis in women.
• Deep learning models for mammogram analysis often require detailed annotations (detection/segmentation).
• Existing methods may overlook critical lesion areas when only using image-level labels.

Purpose of the Study:

• To develop a novel deep learning method for automated breast cancer diagnosis in mammography.
• To focus on local lesion areas using only image-level classification labels.
• To avoid the need for precise detection or segmentation annotations.

Main Methods:

• Proposed a novel adaptive convolutional feature descriptor selection (AFDS) structure.
• Utilized the distribution of deep activation maps to select discriminative feature descriptors.
• Employed a triangle threshold strategy to identify critical feature descriptors (local areas).
• Integrated AFDS into existing convolutional neural networks.

Main Results:

• Ablation experiments and visualization confirmed AFDS aids in differentiating malignant from benign/normal lesions.
• The AFDS structure functions as an efficient pooling mechanism.
• The method integrates seamlessly into existing deep learning architectures with minimal overhead.
• Experiments on INbreast and CBIS-DDSM datasets showed satisfactory performance compared to state-of-the-art methods.

Conclusions:

• The proposed AFDS method effectively enhances breast cancer diagnosis from mammograms.
• This approach offers a valuable alternative for automated screening, reducing annotation requirements.
• The method demonstrates strong potential for clinical application in mammography analysis.