VerTE-MT: A Multi-Task Framework with Entropy-Guided Sampling for Vertebrae Segmentation and Localisation in CT

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Automated spinal CT analysis requires accurate vertebrae segmentation and localization.
• Existing methods often use multi-stage pipelines, neglecting the inherent relationship between segmentation and localization.
• Pathological variations and anatomical similarities pose significant challenges to current approaches.

Purpose of the Study:

• To develop a novel single-stage, multi-task learning framework (VerTE-MT) for concurrent vertebrae segmentation and centroid localization.
• To improve the accuracy and robustness of automated spinal CT analysis, particularly for challenging cases.

Main Methods:

• Proposed VerTE-MT framework with a shared volumetric encoder, Vision Transformer bottleneck, and dual decoders.
• Integrated entropy-guided sampling to prioritize under-represented vertebrae.
• Concurrent segmentation and localization tasks within a single network.

Main Results:

• Achieved high Dice scores for vertebral column segmentation (84.18%-85.45%) and L6 segmentation (81.03%-75.96%) on the VerSe'20 dataset.
• Reduced segmentation boundary errors, with Hausdorff Distance decreasing by up to 4.62mm.
• Maintained robust localization with mean error below 10mm.
• Demonstrated zero-shot validation performance on scoliotic datasets (mean Dice 83.03% and 65.17%).

Conclusions:

• VerTE-MT offers a superior single-stage approach for vertebrae segmentation and localization compared to existing methods.
• The framework shows significant potential for handling pathological spinal anatomies and unseen cases.
• This approach enhances the efficiency and accuracy of automated spinal CT analysis.