INTRODUCTION

Quantitative T1ρ mapping has shown to provide promising imaging biomarkers for early detection of biochemical changes in knee cartilage.¹ Magnetization-Prepared Angle-Modulated Partitioned k-Space Spoiled Gradient Echo Snapshots (3D MAPSS) is a fast imaging technique for high resolution quantitative T1ρ mapping. Recent studies suggest that the 3D T1ρ MAPSS sequence has good reproducibility across different MR systems at different sites.² However, the signal from synovial fluid (SF), if not suppressed, can potentially undermine T1ρ quantification, particularly in clinical conditions where SF increases near cartilage. Notably, SF contaminations prevent reliable evaluation of the superficial layer of the cartilage, which is the area of concern in early stage of osteoarthritis (OA). A long-T2-selective inversion (LT2SI) approach has been proposed to eliminate the contamination of SF to knee cartilage T1ρ quantification at the cost of increased scan time by 50%.³ The additional time penalty needed for SF suppression during imaging will increase overall healthcare cost, reduce patient comfort, and potentially compromise quantitative accuracy due to patient motion. The purpose of this study was to use deep learning to remove SF from T1ρ data acquired without active fluid suppression during imaging.

METHODS

Figure 1 illustrates the 3D MAPSS T1ρ MRI sequence without (a) and with (b) the LT2SI module (90°-τ -180°- τ -90°), respectively.⁴ The inversion time (TI) was optimized to null SF signal before the T1ρ preparation module defined as 90°-TSL/2 -180°-TSL/2-90° (TSL: time of spin lock). Eight volunteers were scanned using a 16 channel T/R knee coil on a 3T clinical scanner (Ingenia Elition, Philips Healthcare, The Netherlands). For each subject, 3D MAPSS T1ρ imaging with full coverage of the knee joint in a sagittal view was performed with and without the LT2SI module (see Table 1 for detailed sequence parameters). The study was approved by local institutional review board and patient consent was obtained from all subjects. Two 4D knee datasets (x×y×z×TSL: 350×175×48×4) with and without SF suppression were available for each subject. The preprocessing step included co-registration of the 3D volumes across the TSL dimension and resize of the datasets to 256×256×48×4. Figure 2 shows the workflow for using deep learning to obtain SF suppressed T1ρ maps from scans without LT2SI. Firstly, two sets of 3D T1ρ maps [x×y×z] were calculated for each subject using conventional non-linear exponential curve fitting. The T1ρ map without SF suppression (T1ρ-nosup) was then combined with the four TSL source images to form a five-channel dataset as input to the deep learning model. The T1ρ map obtained with LT2SI (T1ρ-sup) was used as the target image (ground truth). The deep learning model consisted of an encoder and decoder with skip connections to facilitate feature propagation. The loss function was defined as mean squared error (MSE) = $$$ \frac{1}{n}\sum_{i=1}^n(y_{i}-\widetilde{y}_{i})$$$, where $$$y_{i}$$$ and $$$\widetilde{y}_{i}$$$ are T1ρ maps of the target and model prediction. Table 2 shows the parameter settings of the deep learning model. Datasets of seven subjects were used for training (80%) and validation (20%), and the remaining subject was used as unseen dataset for testing. Normalized mean squared error (NMSE) was used to evaluate the model performance on the test dataset. All deep learning experiments were performed on a workstation equipped with NVIDIA GTX 1080Ti GPU, Intel Core i7-8700K CPU, and 32GB RAM.

RESULTS

Figure 3 shows example images of T1ρ-nosup (a) and T1ρ-sup (c) from conventional curve fitting method, along with the T1ρ map predicted by the deep learning model (b). Distinctively high T1ρ values from SF can be easily identified near the knee cartilage on the T1ρ-nosup map (a), while they were much reduced on the T1ρ maps of the T1ρ-sup (c) and deep learning prediction (b). Pairwise difference images (d-f) were calculated to further illustrate the performance of deep learning model in suppressing SF (white arrows). For the test subject, the NMSE was reduced from 0.1310 (T1ρ-nosup) to 0.1140 (model prediction), using the target (T1ρ-sup) as reference images.

DISCUSSION

Accurate measurement of T1ρ values in knee cartilage is critical for quantitative assessment of early changes in patients with OA. The T1ρ method must have high sensitivity and accuracy to detect a change of T1ρ values at the level of 11-14% to suggest a true change in cartilage composition.⁵ It is well known that OA is usually accompanied by increased SF with T1ρ value 10 times larger than that of the cartilage.⁶ If not suppressed, the high T1ρ value of SF may reduce the quantification accuracy of knee cartilage, particularly for voxels at the superficial layer of the cartilage due to partial volume effect and spatial blurring. The LT2SL module has shown to effectively suppress SF on T1ρ maps, but at the cost of increased scan time by 50%. This study demonstrated the feasibility of using deep learning to remove SF from T1ρ-nosup images without additional scan time. Further studies are warranted to investigate the influence of SF on T1ρ quantification in different cartilage segmentations in a larger patient cohort.

CONCLUSION

Deep learning method can be used to obtain synovial fluid suppressed 3D T1ρ mapping of knee cartilage without adding scan time. This will help improve T1ρ quantification accuracy and facilitate its clinical adoption.

Acknowledgements

None.