Exploiting Patch Sizes and Resolutions for Multi-Scale Deep Learning in Mammogram Image Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning (DL) shows promise for breast cancer screening via computer-aided detection (CAD) systems.
• Patch-based DL methods for 2D mammogram classification face limitations due to fixed patch sizes and unclear resolution impacts.
• Optimal patch size and image resolution for detecting diverse breast cancer lesion sizes remain challenging.

Purpose of the Study:

• To investigate the influence of patch size and image resolution on DL classifier performance for 2D mammograms.
• To develop and evaluate novel multi-scale classifier architectures that integrate diverse patch sizes and resolutions.
• To enhance the accuracy of DL-based CAD systems for breast cancer screening.

Main Methods:

• Proposed multi patch-size and multi-resolution DL classifier architectures.
• Implemented multi-scale classification by combining varied patch sizes and input image resolutions.
• Evaluated classifier performance on the public CBIS-DDSM dataset and an internal dataset.

Main Results:

• The multi-scale classifier achieved an Area Under the Curve (AUC) of 0.809 on the CBIS-DDSM dataset, a 3% improvement.
• On an internal dataset, the multi-scale classifier reached an AUC of 0.722, a 5% improvement.
• The proposed multi-scale approach outperformed baseline single patch size and single resolution classifiers.

Conclusions:

• Multi-scale classification effectively addresses the limitations of fixed patch sizes and single resolutions in mammogram analysis.
• Integrating multiple patch sizes and resolutions significantly enhances the performance of DL-based CAD systems for breast cancer screening.
• The developed multi-scale architectures offer a promising advancement for more accurate and reliable mammogram interpretation.