The application of transfer learning to T1ρ MRI tibiofemoral cartilage segmentation

Proteoglycans are critical to the ability of cartilage to withstand loading, and their concentration can be probed in vivo through measurement of T1ρ MRI relaxation times. Machine learning models have demonstrated the potential to replace manual segmentation of MR images but are often tailored to a specific MRI sequence. Promisingly, transfer learning provides an avenue to leverage information learned by previous models to improve new model performance. Thus, the purpose of this work was to evaluate the utility of transfer learning with source models trained for the same segmentation task (tibiofemoral cartilage) but different MRI sequence (double-echo steady-state vs. T1ρ). For model training, a U-Net architecture, a dice similarity coefficient (DSC) loss function, and a hyperparameter optimization grid search were utilized. Seventy-four T1ρ scans were used to develop and evaluate the models (training: 42, validation: 16, testing: 16). Finally, the optimal transfer learning and non-transfer learning models were applied to the testing set, where model-generated segmentations and T1ρ relaxation times were compared to manually segmented scans. Transfer learning improved DSCs for both the tibial (+1.2 %) and femoral (+0.8 %) cartilage. Furthermore, transfer learning improved the models' ability to measure T1ρ relaxation times, lowering the mean difference relative to manual segmentation from 1.0 % to 0.1 % for tibial cartilage (ICC increased from 0.93 to 0.95) and from 0.6 % to 0.5 % for femoral cartilage (ICC remained at 0.99). These results demonstrate the benefit of utilizing this transfer learning approach and justify using these models in future studies evaluating tibiofemoral cartilage proteoglycan content.