Exploring viability of Test-Time Training: Application to 3D segmentation in Multiple Sclerosis

Test-Time Training (TTT) is an unsupervised domain adaptation technique employing a self-supervised task performed by an attached branch model. While justification of key design choices are often neglected in the literature, we explore the viability of TTT in a real-case scenario and conduct an extensive evaluation of key hyperparameters in TTT, including the choice and the placement of the auxiliary task, the type of normalization layer, and the model parameters to adapt. We carry out this study in Multiple Sclerosis (MS) diagnosis relying on focal lesions visible in conventional MRI. Manual lesion segmentation based on automated methods face significant challenges in clinical integration due to MRI domain shifts, i.e. discrepancies between training and deployment data, notably due to different acquisition settings. We apply TTT to each patient individually, offering an effective strategy to mitigate domain shift without the need of additional annotated data or data from other patients. We ground our experiments on real-world distribution shifts using three distinct MS datasets. Finally, we propose general guidelines to apply TTT in practice. Code available at github.com/Gerin-Benoit/ttt-for-multi-sclerosis.