This paper proposes a novel interactive segmentation model, MOIS-SAM2, for efficient segmentation of neurofibromas (NFs) in whole-body magnetic resonance imaging (WB-MRI) images of patients with neurofibromatosis type 1 (NF1). MOIS-SAM2 extends the state-of-the-art transformer-based promptable segmentation Nothing Model 2 (SAM2) with example-based semantic propagation. It is trained and evaluated on a dataset of 119 WB-MRI scans from 84 NF1 patients, divided into four test sets reflecting within-domain and domain-shifting scenarios (variable MRI field strength, low tumor burden, and differences in clinical site and scanner manufacturer). On the within-domain test set, MOIS-SAM2 achieves a scan-by-scan DSC of 0.60 against expert manual annotation, outperforming the baseline 3D nnU-Net (DSC: 0.54) and SAM2 (DSC: 0.35). It maintains excellent performance across various domain-shifting scenarios, particularly in the case of low tumor burden. The F1 scores for lesion detection ranged from 0.62 to 0.78 across the test set. The agreement between the model and the expert was comparable to that between experts. In conclusion, MOIS-SAM2 demonstrates its potential for integration into clinical workflows by enabling efficient and scalable NF segmentation with minimal user input.
