SIMULTANEOUS 3D CARTILAGE T2 MAPPING AND MORPHOLOGICAL IMAGING WITH RAFO-4 MRI, A MACHINE LEARNING ALGORITHM

INTRODUCTION

Cartilage T₂ is a non-invasive, microstructural MRI biomarker for KOA, with elevated T₂ indicating early KOA onset. Cartilage T₂ maps could be used in clinical trials to test a drug candidate’s effect on microstructure. Quantitative DESS (qDESS) is widely used for cartilage imaging as it simultaneously acquires 3D, morphological whole knee images and quantitative T₂ maps in ∼5 minutes. Researchers are also developing T₂ mapping techniques using phase-cycled balanced Steady State Free Precession (pc-bSSFP). It is rapid and has higher SNR efficiency than qDESS, which could lead to better 3D morphological image quality and more reliable T₂ maps. PLANET is a technique that uses a minimum of six different pc-bSSFP acquisitions to analytically calculate T₂. This is too time-consuming to be clinically feasible. In this study, we trained Random Forest (RaFo) machine learning models to estimate T₂ from fewer pc-bSSFP acquisitions to reduce scan time while still estimating reliable voxel-level T₂ values.

OBJECTIVE

1) Train and test RaFo models on simulated 4 and 6 pc-bSSFP data and benchmark performance with PLANET. 2) Test RaFo models on in vivo knee data and benchmark performance with the reference T₂ mapping technique (spin echo), PLANET, and qDESS.

METHODS

70,000-sample training and 30,000-sample testing datasets were simulated. Each sample corresponded to 12 different pc-bSSFP measurements of the same voxel location in the tissue. The physics-informed simulated datasets were pre-processed, which included sub-sampling from 12 pc-bSSFP measurements to 4 or 6. RaFo models were then trained to estimate T₂ and tested on these pre-processed datasets. Finally, to evaluate performance on noisier in vivo data, fully sampled knee images of two healthy volunteers (HVs, 2F:24-25) were acquired on a 3T Siemens Verio (Erlangen, Germany) with an 8-channel knee coil using 12 measurements of bSSFP (water excitation, 8.6/4.3 ms TR/TE; 22° flip angle; 1 × 1 × 5 mm³ voxel volume; 128 × 128 × 130 mm³), qDESS (water excitation; 20° flip angle; 21.77 ms TR; 6 ms TE; 364 Hz/Px receiver bandwidth; 0 dummy scans per volume), and a gold-standard spin-echo T₂ mapping approach (2500 ms TR; 15, 45, 75 ms TE, 90° and 180° flip angle) with appropriate ethics approval. All images had 1 × 1 × 5 mm³ voxel volume and 128 × 128 mm² field of view. PLANET was tested on 6 pc-bSSFP measurements (labelled PLANET-6). RaFo models were tested on 4 and 6 bSSFP measurements (labelled RaFo-4 and RaFo-6, respectively).

RESULTS

Fig1 shows results from simulated data tests, with similar performance across the RaFo models and PLANET. Fig2 shows the in vivo T₂ maps, with the RaFo models visually aligning best with the reference T₂ maps while qDESS is biased towards lower values in HV1 and PLANET estimates large-valued outliers in HV2 (not visualized). The RaFo models had lower 95% confidence intervals of the difference between the reference and estimated T₂ (∼36ms) compared to qDESS (∼49ms) and PLANET (∼275ms).

CONCLUSION

The RaFo models best aligned with the reference T₂ maps, even when estimating T₂ using only 4 pc-bSSFP acquisitions. They also only estimated biologically feasible values as it can only estimate T₂ values it was trained on, a unique feature of the RaFo algorithm. Hence, RaFo-4 is a promising alternative to qDESS for cartilage morphological and quantitative imaging as it has the potential to have comparable scan times to qDESS, provide better morphological images and estimate more reliable T₂ maps. Future work includes testing RaFo-4 and qDESS on a larger cohort of early KOA patients and HVs.