AI model development for breast cancer treatment and outcome management
We are developing state-of-the-art deep learning models using mammogram imaging and clinical data.
Longitudinal quantification of BAC in 5 patients.
A developed deep learning method for segmenting and quantifying breast arterial calcifications has been developed that can be applied retrospectively to routine screening mammograms. This will allow for the analysis of large populations without additional imaging costs or radiation exposure. Future studies will determine the performance of this tool for predicting clinical outcomes and determining the efficacy of prevention approaches.
We created five novel quantification metrics that assess the segmentation area and pixel intensities in segmented masks for comparison to the ground truth:
1) Sum of Mask Probability (MP): the sum of all pixel probabilities for the whole image;
2) Mask Pixel Count (PC): the total pixel count in the predicted mask;
3) Sum of Pixel Intensity (PI): the sum of all pixel intensities in the predicted mask;
4) Calcified Pixel Count (CPC): the total number of pixels with intensity > 100 in the predicted mask;
5) Calcified Pixel Intensity (CPI): the sum of all pixel intensities >100 in the predicted mask.
Quantification metric correlation (QMC) is defined as the correlation of quantification between the predicted mask and the ground truth mask. QMC was high across all metrics, with CPC slightly outperforming others with an R2 coefficient of 0.973. We compared quantification metrics to calcified volume (voxels) and calcium mass (mg) as measured by breast computed tomography in 12 subjects from prior work 36. Quantification metrics demonstrated R2 coefficient values of 0.798 – 0.843 compared to calcium volume and 0.774 – .873 compared to calcium mass (mg).