Abstract Details

MRI lesion segmentation is a critical step for precision monitoring of multiple sclerosis (MS). Recently, AI-enhanced MS lesion analysis has shown promising performance with fully automatic segmentations. However, the deployment of off-the-shelf AI models trained with non-locally acquired data suffers from performance deterioration due to unseen data features in the inferencing.

To optimize local adaptation for Artificial Intelligence (AI)-based MS lesion segmentation methods with transfer learning.

We propose a Domain Adaptive Lesion Segmentation framework (DAMS-Net) trained with the labelled data from one source MRI scanner, and the unlabelled data from 6 other target scanners. Based on 3D U-Net, DAMS-Net further introduced an adversarial learning-based domain discriminator to alleviate the domain bias of imaging data from different scanners. Our experiments were completed with an MS imaging dataset containing 446 FLAIR scans with completed lesion annotations from 7 different scanners. For each target scanner, we randomly split the data as follows: 80% model training, 20% testing. Labelled data from the 6 target scanners were only used to evaluate segmentation performance. Average Dice and voxel-wise Dice scores are reported.

The average and voxel-wise Dice scores for direct deployment of the pre-trained model on one labelled scanner to other unlabelled scanners were 0.4295 and 0.5275 respectively.  With implementation of DAMS-Net, performance was improved to 0.4720 and 0.5765, respectively.

We demonstrate a domain adaption technique to boost the performance of AI-based MS lesion segmentation on multiple scanners under a label-efficient setting, i.e. without the need to retrain models with new local data and lesion annotations from each center. DAMS-Net represents an effective measure to improve the robustness for MS lesion quantification across large multi-center clinical trials.