Building Convolutional Neural Networks for Abdominal Aorta-Iliac Calcification Assessment and Cardiovascular Risk Prediction of Prostate Patients Undergoing Stereotactic Radiotherapy
H Lin*, W Chen, J Baal, T Upadhaya, J Barrios Ginart, M Roach, III, J Hong, O Morin, University of California San Francisco, San Francisco, CA
Presentations
PO-GePV-T-116 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: We built Convolutional neural networks (ConvNets) to automatically segment the Aorta in non-contrast pelvic CT and quantify the calcification burden to screen Major Adverse Cardiovascular Events (MACE) among prostate SBRT patients.
Methods: A single-institution retrospective study was conducted of 146 patients who underwent SBRT for localized prostate cancer. Fifty non-contrast Pelvic CT scans were used to train two-stage concatenated 3D U-Nets: the localization U-Net used dilated convolution windows to capture location information from large receptive fields, the segmentation U-Net took in two-channel volumes - the bounding box of Aorta created by the localization U-Net and the original CT images, to perform the Aorta auto-segmentation. Twenty auto-segmented Aorta contours were compared with physician contours using Dice similarity coefficient and the average Hausdorff distance. The model was applied to create Aorta contours of the entire dataset. Quantitative calcification burden (CaQ) in units of mm3 was created using a per-voxel threshold of 130 Hounsfield Units. MACE was collected from medical chart reviews. Calcification Burden Score (CaBS), a pre-defined 5-point scale, was derived from the CT review by two physicians who were blinded to outcomes.
Results: Between expert's and auto-segmented Aorta contours, the Dice and average Hausdorff distance were 0.87±0.08 and 1.9±0.6mm. Among the cohort, 20 patients experienced MACE, 49 patients were scored severe CaBS of 3 or higher. CaBS was found positively correlated with MACE (P<0.0001). CaQ derived from Aorta auto-segmentation was highly correlated with CaBS (Pearson r=0.79) and achieved a similar area under the curve (AUC 0.74 vs 0.73 by CaBS) for discriminating MACE, using a cut-off of 1650 mm^3.
Conclusion: A ConvNets-based calcification burden evaluator showed high correlation with expert quantification to identify MACE in prostate SBRT patients, and is promising to address the unmet need of screening and mitigation of cardiovascular risks of radiotherapy patients to improve long-term survival.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email