A POST-PROCESSING METHOD FOR REFINEMENT OF CT-BASED DEEP LEARNING ACETABULAR SEGMENTATIONS.

Authors

Abstract

Accurate acetabular segmentation is important for clinical evaluation of hip morphology and surgical planning. TotalSegmentator (a neural network-based method) enables scalable CT-based hip bone segmentation, but focal errors persist in the acetabular region despite strong overall performance. These limitations highlight the need for targeted regional refinement of the acetabulum. To develop and validate a post-processing refinement method to improve CT-based acetabular segmentations generated by TotalSegmentator. A post-processing workflow was developed in Python to refine TotalSegmentator-generated acetabular segmentations. The acetabulum was localized using connected-components analysis of the obturator foramen and a solidity-based rule to identify the acetabular roof. Binary thresholding at 150 HU suppressed over-segmented non-osseous voxels, while HU-guided mask expansion restored under-segmented osseous voxels. Morphological closing, flood-filling, and selection of the largest 3D connected-component were used to close acetabular contour gaps and to remove residual irregularities. Rule-based slice-wise voxel count checks were used to selectively apply larger-kernel fallback morphological closing when closing using the default kernel size was insufficient. The original segmentation was restored when refinement introduced more error. An established subset of the TotalSegmentator dataset comprising 203 hips from 102 CT scans was used for validation. After exclusion of hips with incomplete view, metal implants, focal acetabular lesions, or pelvic fractures, 104 hips from 58 CT scans were included in the final analysis. The original TotalSegmentator hip bone annotations were manually revised to generate corrected ground truth segmentations. Performance of the post-processing workflow was evaluated by comparing post-processing refined and original acetabular segmentations against corrected ground truths using Dice similarity coefficient (DSC) to quantify overlap, and surface distance metrics, including: surface Dice, average surface distance (ASD), and Hausdorff distance (HD; 95th and 100th percentiles) to quantify boundary agreement. One-sided paired t-tests were used to assess improvement. Within the validation dataset, patients had a mean age of 67.3±11.9 years; 63.8% were male, 96.6% of scans were acquired on Siemens scanners, and the median tube voltage was 100 (interquartile range: 100, 120) kVp. Quantitative segmentation performance within the acetabular region is summarized in Table 1. Compared with the original TotalSegmentator output, the refined segmentations showed significantly higher overall DSC and surface Dice scores, as well as significantly lower ASD and HD95, for both left and right hips. HD100 did not differ significantly between the original and refined segmentations for either side. Representative visual examples of acetabular segmentation before and after post-processing refinement are shown in Figure 1. The proposed post-processing method significantly improved the performance of TotalSegmentator-generated hip bone segmentations within the acetabular region, particularly for surface-based measures. By improving acetabular segmentation quality, this refinement workflow may support more reliable downstream hip morphology and joint space analyses. The corrected ground truth segmentations (https://zenodo.org/records/18853791) and post-processing workflow implementation (https://github.com/DrAKLabs/acetabular-segmentation-postprocessing) are publicly available.

Tags

Topics