GTV segmentation in MRI guided radiotherapy with promptable foundation models

Area of Science:

• Medical imaging
• Artificial intelligence
• Radiotherapy

Background:

• Accurate delineation of gross tumor volumes (GTV) on daily MRI is crucial for MRI-guided radiotherapy.
• Specialized models exist but are often tumor-specific.
• Promptable foundation models offer a potential alternative for versatile GTV segmentation.

Purpose of the Study:

• To investigate the efficacy of promptable foundation models for GTV segmentation in MRI-guided radiotherapy.
• To compare the performance of different promptable models and prompt types across various tumor sites.
• To evaluate the potential of these models as an alternative to domain-specific segmentation approaches.

Main Methods:

• Promptable foundation models (SAM2, MedSAM2, nnInteractive) were evaluated using six sparse geometric prompt types (points, boxes, 2D masks).
• A multi-institutional dataset of clinical GTV masks from diverse anatomical sites (abdomen, lung, liver, pancreas, pelvis) on MRI scans was utilized.
• Model performance was assessed using metrics including the Dice Similarity Coefficient (DSC).

Main Results:

• Promptable models generated segmentation masks comparable or superior to domain-specific models, with median DSCs up to 0.85.
• nnInteractive and MedSAM2 demonstrated superior performance (median DSCs 0.75 and 0.70, respectively) compared to SAM2 (0.54).
• Prompts with greater spatial information generally improved results, though this effect was less pronounced for nnInteractive and MedSAM2.

Conclusions:

• Promptable foundation models hold potential for GTV segmentation in MRI across multiple tumor types.
• Further research is needed to enhance model performance and reduce output variance for clinical implementation.
• These models represent a promising avenue for improving efficiency and accuracy in radiotherapy planning.