ADAPT-SEG: A MODALITY-AGNOSTIC PIPELINE FOR ROBUST PATIENT-SPECIFIC MSK SEGMENTATION

Abstract

INTRODUCTION

Accurate segmentation of musculoskeletal MRI is a critical step in patient-specific musculoskeletal modeling and quantitative analysis of cartilage. However, segmentation remains a time-consuming bottleneck limiting the feasibility of large-scale studies. While machine learning tools can automate segmentation in large, homogeneous datasets, they often struggle to generalize to unfamiliar image types, poor-quality images, or extreme joint pathology. Augmenting machine learning segmentation with modality-independent statistical shape models (SSMs) could improve generalization by constraining outputs to fit reasonable shape priors, thereby reducing the manual segmentation burden in small and diverse datasets.

OBJECTIVE

We hypothesize that augmenting deep learning segmentation (nnUNet) with modality-independent shape priors will improve segmentation accuracy in unfamiliar MRI sequences compared with nnUNet alone.

METHODS

We developed ADAPT-seg (Agnostic Deep-learning and Prior-guided Tissue Segmentation), a three-step pipeline consisting of: 1) mutual-information based registration to a reference image, 2) initial segmentation boundary estimation with Active Shape Modeling (ASM), and 3) refinement with nnUNet (Figure 1). Using labeled DESS-sequence knee images from the OAI-ZIB dataset (n=40, KL grade 0), we created a 3D SSM of the femur and tibia and an nnUNet “fullres-cascade” two-stage segmentation model (Figure 1). Five additional KLG 0 participants with DESS, FLASH and TSE-sequence images comprised the test set. All test images were segmented using nnUNet alone and the full ADAPT-seg pipeline, with ASM-derived labels serving as the low-resolution stage. Performance against manual segmentations was evaluated in a region of interest (ROI) covering the key articulating anatomy using Dice scores, average surface distances, and boundary F1 scores.

RESULTS

ADAPT-seg demonstrated greatly improved segmentation accuracy on unfamiliar FLASH and TSE sequences in all metrics and performed comparably to nnUNet on DESS images (Table 1, Figure 2). nnUNet failed to correctly label the tibia in all TSE and FLASH images, a catastrophic failure prevented by SSM-based initialization. Both nnUNet and ADAPT-seg added tibia labels to the patella in all FLASH scans.

CONCLUSION

SSM-based augmentation is a feasible strategy to improve generalization of nnUNet segmentation across unfamiliar MRI sequences. ADAPT-seg remains under active development, with the aim of extending the pipeline to cover other joints, more extreme pathomorphology, and multiple tissue types including articular cartilage. Our long-term goal is to develop ADAPT-seg into an open-source tool which, used alongside expert review and correction, can greatly reduce the manual segmentation burden for MSK researchers.