Performance of an Artificial Intelligence Foundation Model for Prostate Radiotherapy Segmentation

Radiation therapy is a major treatment modality for localized prostate cancer, and accurate target segmentation is a critical aspect of radiation delivery that can directly affect patient outcomes. We evaluated the performance of an artificial intelligence (AI) foundation model in target segmentation tasks for prostate radiotherapy planning.

AI segmentation using Segment Anything Model 2 (SAM 2) was evaluated from computed tomography images with varying levels of human intervention, ranging from intervals of every 2nd to every 10th ground truth slice provided as input. Segmentation accuracy was measured using dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95) for intact and postoperative prostate target delineation.

SAM 2 achieved performance similar to and, in some cases, worse than interpolation in DSC and HD95 for both intact and postoperative prostate cancer patient cases, and the AI segmentation accuracy was significantly better in the intact preoperative patient cases where anatomic boundaries were better defined than postoperative patient cases (P < .01† for intervals 5-10). When sparse ground truth was provided simulating lower levels of human intervention, DSC values decreased from 0.956 to 0.856 in the preoperative setting and from 0.952 to 0.751 in the postoperative setting when the interval between ground truth slices was increased from every 2nd to every 10th slice. Correspondingly, the HD95 increased from 1.817 to 4.436 in the preoperative setting and from 1.988 to 11.02 in the postoperative setting under the same conditions.

Current general-purpose AI foundation models do not outperform existing methods and are likely inadequate for prostate radiotherapy segmentation. Their clinical application will require further understanding of appropriate fine-tuning and task-specific performance.