Detailed Segmentation of Pelvic Arteries in Pelvic CT Angiography with Deep Learning

Atherosclerosis of the pelvic arteries leads to diseases such as erectile dysfunction{1}, lower urinary tract symptoms and buttock claudication{2}. Understanding the anatomy of pelvic arteries is important for the undertaking of endovascular intervention. CT angiography is the main modality for initial evaluation and has high sensitivity{3}. Detailed and accurate segmentation of pelvic arteries in CT angiography is the cornerstone for the automated diagnosis and quantification of pelvic arterial diseases. However, no detailed segmentation model for pelvic CT angiography was described in the literature to date.

Materials and Methods:

A dataset containing 50 pelvic CT angiography studies from 50 male patients was collected. The patients’ mean age was 63.1, and the mean number of slices was 516.3. Each slice contains 512×512 voxels. The spacing was 0.423 mm, and the thickness was 0.625 mm. The segmentation masks were manually annotated, including only vessels branching distally from the aortic bifurcation. The train, test, validation splits were 60%, 20% and 20%. We used the dense UNet architecture for segmentation{4}. A combination of the region-based Tversky loss and the contour-based boundary loss was used for optimization{5}. We additionally built a two-model pipeline with a global model predicting the entire study and a local model focusing on voxels inside the pelvic cavity to enhance small vessel segmentation performance. Then this combined result underwent automated connected component analysis to remove small fragments, such as renal tissue and mesenteric arteries that are not connected to the desired vascular tree. Each step was finally evaluated by Dice similarity coefficient (DSC) on the test set.

Results:

The best DSCs for the single global model, the two-model pipeline and the connected component analysis were 0.839, 0.844 and 0.862. Significant improvements on DSC were found between the single global model and the two-model pipeline (p=0.003) as well as the two-model pipeline and connected component analysis (p=0.02).

Conclusion: We presented the first deep learning system dedicated for detailed segmentation of pelvic arteries. A single global model achieved a DSC of 0.839. Our two-model pipeline enhanced the DSC to 0.844. After connected component analysis, the DSC was further improved to 0.862.