Deep Vision for Breast Cancer Classification and Segmentation.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Diagnostic Support
• Mammography Image Analysis

Background:

• Rising breast cancer diagnosis rates necessitate improved screening accuracy.
• High false positive rates (FPR) in mammography lead to overdiagnosis and overtreatment.
• False negative rates (FNR) contribute to increased patient morbidity and mortality.

Purpose of the Study:

• To develop and evaluate a deep vision supervised learning model for mammography classification.
• To assess the model's accuracy, specificity, and sensitivity in detecting breast cancer.
• To explore gradient-based techniques for automated region of interest (ROI) identification.

Main Methods:

• Supervised learning models trained on 55,890 de-noised mammography images (299x299 pixels).
• Application of trained models to a test set of 15,364 unseen images for classification.
• Utilizing loss function gradients to identify image regions crucial for classification decisions.

Main Results:

• Achieved 97% accuracy, 99% specificity, and 83% sensitivity in initial classification.
• Gradient techniques successfully mapped influential regions on positive mammograms.
• Demonstrated potential for supporting clinical analysis through automated ROI identification.

Conclusions:

• Deep vision techniques show promise in mitigating mammography overdiagnosis and underdiagnosis.
• Automated ROI identification can assist clinicians in mammogram interpretation.
• AI-powered analysis offers a pathway to more accurate and efficient breast cancer screening.