Bridging field strengths: fine-tuned deep learning models for 7T MRI white matter lesion segmentation in multiple...

Multiple sclerosis (MS) white matter lesion (WML) automated segmentation on ultra-high-field 7T MRI remains challenging due to the domain shift from lower-field acquisitions, limited annotated data, and specific imaging artifacts. Fine-tuning is an effective and practical strategy for adapting deep learning WML segmentation algorithms to 7T MRI, even with limited annotated data. This study evaluates fine-tuning as a domain adaptation strategy to leverage a pre-trained deep learning model for automated WML segmentation on 7T MRI. We fine-tuned a U-Net-based model, originally trained on approximately 35,000 heterogeneous lower-field (1T, 1.5T, 3T) multi-contrast MRI scans of people with MS for T2-hyperintense WML (T2-WML) segmentation, using a 7T dataset. Multiple approaches were evaluated, including standard fine-tuning on 3D FLAIR images, low-rank adaptation (LoRA) and training from scratch (nnU-Net). Models were evaluated on an external multi-center 7T test dataset. Additionally, a separate model was fine-tuned for T1-hypointense WML (T1-WML) segmentation on 7T MP2RAGE images. The original model showed substantial performance degradation on 7T data compared to 3T (Dice score decreased from 0.69 to 0.31), confirming the need for domain adaptation. Fine-tuning markedly improved T2-WML segmentation, with the fine-tuned model achieving a median Dice score of 0.57. Lesion-wise sensitivity and F1-score were 0.78 and 0.75, respectively, and the lesion volume agreement with manual segmentation was 0.93. For T1-hypointense WML, fine-tuning showed lower lesion detection performance compared to the T2-hyperintense WML segmentation (sensitivity and F1 of 0.58 and 0.50, respectively); however, incorporating multi-center data into the training set substantially reduced false positives by 23% and improved lesion detection by 14%. Multi-center fine-tuning further improved performance, particularly for the more challenging task of T1-WML segmentation. The models presented here may be included in MS research workflows to facilitate multi-center collaborations with 7T MRI.