IMPROVING ACCURACY AND REPEATABILITY OF T2 MAPPING IN THE OAI DATA THROUGH EXTEND PHASE GRAPH MODELING

The Osteoarthritis Initiative was a longitudinal study of osteoarthritis that prospectively collected a trove of imaging data including Multi-Echo Spin-Echo (MESE) data for cartilage T2 relaxation time assessment in one knee. While this data remains underutilized, several analyses have been performed over the past years to assess T2 sensitivity to OA exploiting the OAI dataset. However, fitting procedures to compute T2 maps from the MESE data in the OAI largely rely on mono-exponential modelling, which is inherently sub-optimal as it does not account for stimulated echoes produced by RF slice-profile and B1 inhomogeneities and it often fails to account for low SNR in longer TEs. To mitigate errors, a common practice is to drop the first echo and fit the remaining 6 echoes at the expense of discarding information and degrading SNR efficiency. T2 fitting of MESE data using Extended phase graph (EPG) modelling, whether based on nonlinear least square (NLSQ) dictionary matching (DM) or deep learning (DL), can account for stimulated echoes, and can potentially provide more accurate and robust fitting for T2 mapping in the OAI. This work proposes to 1) set up three EPG fitting approaches for T2 mapping in the OAI dataset (NLSQ-based, DM-based and DL-based), 2) assess methods for their performance in accuracy and robustness to noise using both simulations and in-vivo data, and 3) compare them against standard fitting methods based on mono-exponential methods. MESE simulations were performed in Matlab (R2022b) using the EPG formalism considering the sequence parameters of OAI data. Hanning-windowed Sinc pulses were used for slice-profile simulations. Three EPG-based fitting methods and three exponential (EXP)-based methods used in prior OAI literature were considered and are summarized in Figure 1. To investigate fitting accuracy and repeatability robustness to noise experiments with simulations as well as using in-vivo data from OAI database were performed. 2000 MESE signals were simulated with T2 ranging from 20 to 80 ms and B1 ranging from 0.9 to 1.1. Each method was used to fit T2 values after adding increasing levels of Gaussian noise. For each SNR, the procedure was repeated 10 times with re-sampling of noise. Accuracy was assessed using the mean percentage error (MPE) and mean absolute percentage error (MAPE), while repeatability was assessed with coefficient of variation (CV). MESE data from 5 subjects in the OAI database (1 in each KLG) were corrupted by injecting Gaussian noise to the MESE images twice with increasing variance. Method repeatability was assessed through Bland-Altman (BA) analysis. To assess agreement among fitting methods and how this affected inference of the presence of OA, 50 subjects were randomly selected from the OAI dataset: 10 subjects (5F & 5M) per KLG (0,1,2,3,4). Patellar (P) and Tibiofemoral (TF) cartilage T2 maps were computed pixel-wise with all the described fitting methods. Mean T2 was computed in 7 ROIs (P, MT, LT, central and posterior regions for the MF and LF) extracted using automatic segmentation of DESS images registered to MESE images. BA analysis was used to asses pair-wise agreement in mean T2 values using Limits of Agreement (LOA) and mean bias. The Lin's concordance coefficient (ρc) and CV were also used as metrics of agreement. A logistic regression model was then performed using OA presence (KLG≥2) as a dependent variable, T2 as independent variable and body mass index as covariate in the MT and the central MF regions. MPE, and CV for different fitting methods from the simulation experiment are reported in Fig. 2 (top panel) as a function of SNR. The EPG methods outperformed the exponential-based methods in terms of accuracy at all SNR levels. The EPG-DL approach had the best overall performance in terms of accuracy and repeatability. In-vivo analysis of LOA and CV as function of SNR (Fig. 2, bottom panel) showed that the EPG-based methods had higher repeatability than EXP-based procedures. The EPG-DL approach also had the best overall performance in in vivo data. T2 pair-wise method comparison in-vivo (Fig. 3) showed that overall, the EPG-based methods had higher inter-method agreement (- 0.1 ms < Bias < 0.05 ms, 0.2 < LOA < 1.13 ms, ρc ∼ 0.99) compared to exponential-based methods (-0.7 ms < Bias < 2 ms, 3.2 ms < LOA < 5.3 ms, 0.86 < ρc < 0.94). Poor agreement was found between EPG-based and exponential-based methods (0.34 < ρc < 0.44, Bias ∼ 10 ms and LOA ∼ 4 ms). With reference to Tab. 1, using the EPG-based methods resulted in higher T2-associated OA odd ratios than EXP-based methods in the MT region (EPG OR ∼ 1.19, 1.13 < EXP OR < 1.18). EPG-based T2 relaxation time fitting methods resulted in more accurate and repeatable T2 estimation than EXP-based approaches in simulations. Preliminary in-vivo experiments also suggest higher robustness to noise of EPG methods compared to EXP-based methods. Furthermore, the EPG-methods showed high inter-method agreement. The lower T2 inter-method agreement of EXP-based approaches greatly affected inference of OA severity. Despite the limited sample size, these results suggest that EPG-based methods to compute T2 maps in the OAI may result in low method-dependent variability. Among the EPG-based methods, the DL approach showed the highest repeatability. The high repeatability of EPG-DL paired with its computational efficiency may allow better exploitation of T2 information in the OAI dataset, especially when longitudinal analysis is involved. We plan to use the EPG-DL approach to compute T2 maps of the entire OAI dataset and make it publicly available for researchers to use it.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.