An Orientation- and Paired-Edge-Aware Modified Frangi Filter for Semi-Automated Crural Fascia Segmentation in Ultrasound Images of Patients with Stroke

Quantitative assessment of the crural fascia using ultrasonography may provide clinically relevant information on peripheral musculoskeletal changes in patients with chronic stroke. However, accurate segmentation of this thin structure is challenging because of speckle noise, discontinuous boundaries, and competing hyperechoic structures. This study aimed to develop and evaluate a training-label-free, semi-automatic framework for continuous crural fascia segmentation in B-mode ultrasound images. The framework requires manual selection of a region of interest and placement of one seed within the target fascia. It then integrates fascia-specific image enhancement, seed-guided tracking of a continuous center path, joint detection of the superficial and deep boundaries, and continuity-preserving refinement to generate a binary fascial mask. Performance was evaluated using all 80 ultrasound images obtained from 40 participants and expert-generated reference masks. Conventional Frangi filtering and component-wise ablation configurations were used as comparators. The complete framework achieved an intersection over union of 0.812 ± 0.153, centerline Dice of 0.894 ± 0.106, average symmetric surface distance of 0.056 ± 0.041, and 95th-percentile Hausdorff distance of 0.076 ± 0.040. The corresponding values for conventional Frangi filtering were 0.423 ± 0.275, 0.693 ± 0.191, 0.267 ± 0.191, and 0.715 ± 0.518, respectively. These findings indicate that integrating fascia-specific appearance information with anatomical target anchoring and path-continuity constraints can improve continuous crural fascia segmentation. However, because parameter development and performance evaluation used the same dataset, independent validation with prespecified parameters is required to establish generalizability.