Study Population:

We retrospectively analyzed 2D mammograms and CT studies from 1449 women (mean age 60 years +/- 11 [SD]) with known CT examinations of the head, neck, chest, abdomen, or pelvis within 12 months of the mammogram date.

Quantitative BAC Analysis:

A deep learning AI quantitative BAC model was employed, trained on an internal dataset of 2D mammograms. The model detects BAC based on expert radiologist annotation and provides a score of 0-5, considering the total area of BAC and density, where:

     0: no BAC visible

     1: short BAC vessel(s) with faint BAC

     2: short BAC vessel(s) with dense BAC

     3: one long BAC vessel and moderately dense BAC

     4:  one long BAC vessel and dense BAC

     5: multiple long BAC vessels with dense BAC

Fig 1: Fig 1. AI BAC Algorithm: Deep convolution neural network models trained to...

Fig 2: Fig 2. U-Net: Vessel detection/segmentation.

Fig 3: Fig 3. ResNet: BAC vessel classification.

Fig 4: Fig 4. Truth Annotation and Scoring: Expert radiologists annotate BAC vessels...

Visual Assessment of Atherosclerotic Disease:

Two radiologists performed visual assessments of atherosclerotic disease in any large or medium-sized vessel on CT exams. The extent of atherosclerotic disease was categorized from 0 to 5, with moderate or extensive involvement (> 50% by visual assessment) considered clinically significant.

Fig 5: Fig 5. Mammogram and CT abdomen/pelvis performed within 12 months on same...

Fine-Tuning Model with Transfer Learning:

Starting from a pretrained U-Net model using non-medical images, further training of the model was accomplished  using mammograms and visually annotated BAC regions for BAC detection.

Fig 6: Fig 6. Model Training: Transfer learning to fine-tune the model.

Statistical Methods

Estimates of the rate of clinically significant atherosclerotic disease were compared for women with BAC score 0-2 vs. 3-5 using a Fisher exact test of independence.