Glioma Segmentation on Multicenter 2D-Based Magnetic Resonance Imaging Using Low-Rank Adaptation Tuning of a Foundation Model: An External Test Evaluation.

Research on foundation models is actively progressing. The segment anything model (SAM) and MedSAM are representative foundation models for image segmentation. Recently, low-rank adaptation (LoRA) has been developed, allowing parameter updates without retraining the entire model, thus solving the problem with large data and time required for task-specific fine-tuning. Although many studies have used public databases, few have focused on local data. Moreover, to our knowledge, no studies have fine-tuned MedSAM using LoRA. We aimed to evaluate SAM, MedSAM, and their LoRA-tuned variants (SAM-LoRA and MedSAM-LoRA) using brain magnetic resonance images of gliomas from five centers in Japan and to compare their performance. We used 2D-based fluid-attenuated inversion recovery axial images and conducted parameter optimization based on four-fold cross-validation (189 cases) and external test evaluation (75 cases) using cases collected retrospectively. Dice coefficients, intersection over union (IoU), and the 95% Hausdorff distance (HD95) were used to evaluate the performance of SAM, MedSAM, SAM-LoRA, and MedSAM-LoRA. Additionally, subgroup evaluations were performed according to scanner manufacturer, glioma location, and calcification status. In the external test at the case level using SAM-LoRA and MedSAM-LoRA, the Dice coefficients/IoU/HD95 were 0.9166/0.8506/1.3517 and 0.9051/0.8342/1.8061, respectively. Both SAM-LoRA and MedSAM-LoRA demonstrated significantly higher Dice coefficients and IoU, as well as significantly lower HD95 values, compared with SAM and MedSAM. Subgroup evaluations also showed highly accurate extraction across different scanner manufacturers, glioma locations, and calcification statuses. SAM-LoRA and MedSAM-LoRA achieved high accuracy on an externally evaluated dataset, suggesting potential utility.