Towards Robust Supervised Pectoral Muscle Segmentation in Mammography Images.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Mammography is crucial for breast cancer screening, but pectoral muscle presence complicates automated detection.
• Existing pectoral muscle removal methods often rely on traditional machine learning due to a lack of segmentation data.
• Developing robust automated detection systems requires effective pectoral muscle segmentation.

Purpose of the Study:

• To provide pectoral muscle segmentation masks for the INbreast, MIAS, and CBIS-DDSM datasets.
• To facilitate the development of supervised deep learning methods for pectoral muscle removal.
• To evaluate the efficacy of deep learning models trained with segmentation masks for pectoral muscle segmentation.

Main Methods:

• Generated pectoral muscle segmentation masks for INbreast, MIAS, and CBIS-DDSM datasets.
• Trained an AU-Net model for pectoral muscle segmentation on INbreast and CBIS-DDSM datasets.
• Evaluated model performance using cross-dataset testing on unseen data, including the MIAS dataset.

Main Results:

• Pectoral muscle segmentation masks were successfully created for multiple public datasets.
• Deep learning models (AU-Net) demonstrated effectiveness in pectoral muscle segmentation.
• Cross-dataset testing yielded performance comparable to same-dataset experiments, validating generalizability.

Conclusions:

• The provided segmentation masks enable advanced deep learning approaches for pectoral muscle removal in mammography.
• Deep learning models trained with these masks offer a powerful tool for automated breast cancer detection systems.
• Cross-dataset evaluation confirms the robustness and applicability of the developed methods to unseen data.