Semi-Automatic Tumor Segmentation for Response Assessment: RECIST Measurements As Visual Prompts for Nninteractive Model
Abstract
Purpose
State-of-the-art recommendations for tumor response assessment are defined by the Response Evaluation Criteria in Solid Tumors (RECIST), which rely on the tumor axial long and short axes measurements. However, RECIST-based measurements often yield limited information regarding tumor morphology. Volumetric biomarkers promise to provide a more comprehensive assessment of tumor response. However, full volumetric tumor segmentation is not performed routinely as it is technically challenging and time demanding. This study investigated the feasibility of using an AI-based approach with promptable segmentation foundation models to enable fast and accurate volumetric segmentation of lung tumors using RECIST-based measurements.
Methods
A promptable 3D segmentation foundation model (nnInteractive) was utilized to segment lung GTVs using RECIST long and short axis measurements as visual prompts. The study included lung cancer patients (stages I–II) treated with SBRT on 0.35T MRI-Linac systems. Daily MR images from two institutions were analyzed. RECIST long and short axes were used as scribble prompts, combined with negative point prompts derived from the axes. Resulting masks were compared with clinically approved GTVs using standard segmentation metrics and selected radiomic features.
Results
In the internal dataset (122 patients, 583 GTVs), median (IQR) Dice similarity coefficient (DSC), surface DSC, and 95th percentile Hausdorff distance (HD95) were 0.79 (0.67-0.85), 0.85 (0.68-0.94), and 4.64 (3.17-8.51) mm, respectively. Comparable performance was observed in the external dataset (17 patients, 119 GTVs), with median DSC of 0.79 (0.70-0.84), surface DSC of 0.81 (0.73-0.90), and HD95 of 4.72 (3.41-6.22) mm. Median (IQR) relative difference in radiomic features was 0.2 (-4.3-6.6) %.
Conclusion
RECIST-guided prompting of nnInteractive enables accurate and fast volumetric segmentation of lung tumors with minimal user input. This approach has the potential to facilitate routine use of volumetric and radiomic response assessment without increasing physician workload. Ultimately, it promises to support improved tumor response evaluation and clinical decision-making.