Deep Learning Model Enhances Identification of Calcifications on Mammograms

Key Details

• 1The DL model uses a modified U-Net architecture tailored for BAC segmentation and quantification.
• 2In testing with mammograms from 369 women, the model achieved Dice scores of 0.90 (training) and 0.89 (validation) for segmentation.
• 3Classification accuracy for BAC detection was high, with F1 scores of 0.97 (validation) and 0.93 (training).
• 4Bland-Altman analysis validated the model's reliability for calcium quantification, with a mean difference of -0.98 mg in the training set.
• 5Manual annotation by seven radiologists enabled categorization into six BAC mass levels.
• 6The approach could transform mammography into a dual-purpose screening tool for breast cancer and cardiovascular disease risk.