Optimizing Image Preprocessing for Deep Learning-Based Aortic Root Segmentation in TAVR Assessment

Recent advancements in deep learning-based auto-segmentation have enhanced preoperative assessments for transcatheter aortic valve replacement (TAVR), but the impact of image preprocessing on aortic root segmentation remains unclear. This study evaluates the effects of four preprocessing methods— denoising, histogram equalization (HE), resampling, and data augmentation (DA)—on the segmentation of calcified aortic root using four CNN architectures: 3D UNet, Res-UNet, SegResNet, and UNETR. A total of 98 preoperative CTA images were processed with five different pipelines, and segmentation performance was assessed using Dice Similarity Coefficient (DSC), True Positive Rate (TPR), and 3D deviation. Results showed that resampling significantly improved segmentation quality, with DSC and TPR increases of up to 6.07%. Denoising negatively affected calcification segmentation, while HE and DA provided modest improvements. The best performance was achieved with a pipeline combining HE, resampling, and DA, excluding denoising.