Author: Abid Khan, Chad Klochko, Michael J Kovalchick, Hyeok Jun Lee, Hani Nasr, Krishnan Shyamkumar, Kundan S Thind ๐จโ๐ฌ
Affiliation: Henry Ford Radiology, Wayne State University, Henry Ford Health, HFHS ๐
Purpose: Automated vascular segmentation in interventional angiography is challenged by contrast kinetics, vessel variations, and 2D projections, limiting the effectiveness of single-model approaches. This study proposes and validates a two-stage soft segmentation strategy to enhance performance across varying vascular conditions, specifically focusing on vessel width and bifurcations.
Methods: Retrospective review identified 234 patients who experienced interventional fluoroscopy of the celiac axis with iodinated contrast from January 1st 2019 to December 31st 2022. 662 partitions pairs of 128x128 pixels were generated comprising an image and corresponding multi-class mask.Segmentation classes (Trunk, Bifurcation, Periphery), were defined by diameter fraction of the celiac axis and number of vessel bifurcations. These pairs were categorized into four support sets based on class characteristics. A UniverSeg model was uniquely adapted as a first-stage discriminator, generating four softmax output layers representing the probability distribution for each class in the image. These probability maps were compressed and min-max normalized to create a refined input for the second stage, which was split 8:2 for training and validation. A 2D-nnU-net was employed as the refining model, selected for its proven effectiveness in medical image segmentation. Model performance was evaluated using five-fold nested cross-validation.
Results: Mean Balanced-Average-Hausdorff-Distances were 0.46, 0.58, 0.61 (ฯ=0.38, 0.43, 0.99) across the class specific tests, indicating strong spatial agreement in the specialized cases and improvement over comparative nnU-net results (ยต=0.75, 1.16, 3.28; ฯ=0.31, 0.55, 0.94). While variance increased in cases with diverse vessel arrangements (ฯ= 0.84, 1.32, 3.55), mean performance decline was small (ยต = 0.57, 0.62, 0.65) in absolute pixel difference against control conditions. All evaluation metrics increased with vessel diameter, reduced bifurcations and dataset augmentation.
Conclusion: This two-stage soft segmentation model demonstrates promising results for specialized problems within interventional angiographic segmentation, as showcased by its development and testing on this large and diverse dataset.