A SYSTEMATIC POST-PROCESSING APPROACH FOR T1Ρ IMAGING OF KNEE ARTICULAR CARTILAGE

INTRODUCTION

T_1ρ imaging is an emerging technique in knee MRI for the evaluation of OA. This modality possesses the unique capability to image biochemical components, such as proteoglycans, facilitating early detection and post-treatment monitoring of knee OA. However, a significant challenge associated with T_1ρ imaging lies in the complexity of its post-processing, which encompasses parameter fitting, cartilage segmentation, and subregional parcellation.

OBJECTIVE

This abstract presents a systematic methodology for automating knee T_1ρ MRI post-processing by leveraging deep learning and advanced computational techniques.

METHODS

Our methodology automated the three primary steps of T_1ρ knee MRI post-processing and provided the mean T_1ρ values for 20 subregions of the femoral and tibial cartilage in the knee (Figure). In our experiments, we utilized four T_1ρ-weighted images to generate the T_1ρ map for 30 OA patients (age 67.63±5.80 years, BMI 26.00±4.08 kg/m²) and 10 healthy volunteers (age 24.90±2.59 years, BMI 22.75±4.51 kg/m²). For each subject, four T_1ρ-weighted images were acquired using a spin-lock frequency of 300 Hz and spin-lock times of 0, 10, 30, and 50 ms, with a resolution of 0.8 × 1 × 3 mm³, resulting in an image matrix size of 44 × 256 × 256 . The spin-lock preparation was followed by an FSE readout with TE/TR = 31/2000 ms. Additionally, we computed the mean of the four T_1ρ-weighted images and employed this mean for automated cartilage segmentation and subregion parcellation. We employed a nnU-Net trained with all 40 subjects for cartilage segmentation, while subregion parcellation was conducted using our previously published rule-based method, CartiMorph. The performance of the approach using deep learning segmentation was assessed using the Dice Coefficient Similarity (DSC), the root-mean-squared deviation (RMSD), and the coefficient of variance of RMSD (CV_RMSD) against the manual segmentation. We excluded 3 OA patients with full cartilage loss above 50% of one cartilage area (FC, MTC, or LTC) in subregion analysis.

RESULTS

Our experimental results demonstrated the satisfactory performance of our proposed approach. The mean DSC values for the FC, MTC and LTC in OA patients and healthy volunteers were 0.83, 0.80, and 0.82, respectively. Table 2 provides a comprehensive breakdown of the performance metrics of the agreement in T_1ρ quantification across 20 subregions.

CONCLUSION

We proposed a systematic approach for post-processing knee T_1ρ MRI data. The experimental results demonstrated the efficacy of the proposed approach.