AUTOMATING IMAGING BIOMARKER ANALYSIS FOR KNEE OSTEOARTHRITIS USING AN OPEN-SOURCE MRI-BASED DEEP LEARNING PIPELINE

Abstract

INTRODUCTION

Quantitative MRI and [¹⁸F]NaF PET enable assessment of cartilage composition, bone shape, and subchondral bone metabolism in knee OA. Current workflows rely on manual segmentation that is time-consuming and subject to inter- and intra-reader variability. Furthermore, computing quantitative metrics requires considerable time and expertise. An open-source, automated, deep learning (DL) pipeline with standardized biomarker extraction has the potential to enhance reproducibility and make large-scale analysis accessible to clinical research communities, including non-technical users.

OBJECTIVE

Develop and validate an automated DL-based pipeline for comprehensive MRI-based segmentation and quantitative analysis of multiple knee tissues from multi-modal MR and PET images.

METHODS

We developed and open-sourced a comprehensive segmentation and analysis pipeline. A 2D U-Net was trained to segment 9 tissues using a dataset of 347 DESS and qDESS images: 3 bones (femur, tibia, patella), 4 cartilage regions (femoral, medial and lateral tibial, patellar), and 2 menisci (medial and lateral). Subchondral bone masks and femoral cartilage subregions were fitted automatically. Quantitative imaging biomarkers were computed as follows: cartilage T2 was computed analytically from qDESS scans; cartilage thickness was computed as the 3D Euclidean thickness of cartilage overlying the bone surface; meniscal volume was calculated as the product of voxel count and voxel volume; OA bone shape (BScore) was derived using a neural shape model; PET-derived subchondral bone metabolism was computed as regional SUVmean/max, and kinetic modeling via Hawkin’s method was used to extract KiNLR (bone mineralization rate) and K1 (perfusion to subchondral bone). To evaluate the pipeline, 20 unilateral qDESS and [¹⁸F]NaF PET knee scans (10 symptomatic OA, 10 controls) were analyzed by the automated pipeline, and two manual annotators. Manual and automated segmentations were compared using the Dice Similarity Coefficient (DSC) and average symmetric surface distance (ASSD). Biomarkers were compared using ICC and normalized mean RMSE (NRMSE).

RESULTS

All automated segmentations had good to excellent overlap measured using DSC (bone: 0.95-0.98; cartilage: 0.84-0.91; menisci: 0.85-0.89) and small surface errors (bone: 0.13-0.32 mm; cartilage: 0.11-0.21 mm; menisci: 0.17-0.30 mm). Notably, automated segmentations had better DSC and ASSD than the inter-rater comparison (Fig. 2). With the exception of cartilage thickness and patellar cartilage whole T2 values, all quantitative metrics showed excellent agreement with ICC >0.96 and NRMSE <0.1, comparable to inter-rater comparison. Bone metrics (BScore, SUV, PET kinetics) had ICC >0.96. Cartilage metrics had more variability, with the best reproducibility for whole cartilage T2 (ICC 0.89-0.98, NRMSE 0.01-0.04), then superficial T2 (ICC 0.93-0.99, NRMSE 0.01-0.05), and finally deep T2 (ICC 0.7-0.97, NRMSE 0.01-0.06). Cartilage thickness showed the worst reproducibility but still was comparable to inter-rater measures. Meniscus volume also shows high concordance (ICC 0.93-0.97; NRMSE 0.05-0.10). Overall, we found that most of the metrics derived from automated segmentations are comparable to those derived from manual segmentations.

CONCLUSION

Our open-source, AI-driven pipeline delivers rapid, accurate segmentation and quantitative analysis of multimodal knee MRI and PET data. Next steps include support for other MR sequences, multi-site validation, and 3D Slicer integration to facilitate translation. This resource provides a foundation for reproducible and scalable imaging biomarker analysis in OA research and clinical trials.