Author: Jihun Kim, Jin Sung Kim, Jun Won Kim, Yong Tae Kim, Chanwoong Lee, Jihyn Pyo, Young Hun Yoon 👨🔬
Affiliation: Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Department of Radiation Oncology, Gangnam Severance Hospital, Yonsei University College of Medicine 🌍
Purpose: Although segmentation foundation models have recently demonstrated promising zero-shot performance on natural images, its clinical application to magnetic resonance (MR) images still requires additional development for task-specific and streamlined implementation. This study proposes a fully automated zero-shot organ segmentation framework in male pelvic MR images acquired during MR-guided radiation therapy (MRgRT) without fine-tunning of a pretrained model.
Methods:
Data sets of 84 prostate cancer patients collected for this study included planning computed tomography (pCT) and T2 MR images, and structures defined on both imaging modalities (femoral heads, bladder, and anorectum). The proposed segmentation framework was developed based on the Segment Anything Model2 (SAM2), a prompt-based segmentation model. Initial point prompts were generated by performing organ-specific deformable image registrations (DIRs) from pCT to T2 MR images and calculating the center-of-mass (COM) of the deformed structure. For each organ, its COM was used as a foreground-prompt (to be segmented) while those of the other organs served as background-prompts (to be excluded). To improve performance, SAM2-based segmentation was performed in multiple orientations: axial, sagittal, and coronal slices. The results from these multi-directional segmentations were combined to produce the final segmentation outcome. The performance of the proposed method was compared to other methods based on DIR, SAM, SAM in medical images (MedSAM), and TotalSegmentatorV2 (TS). The Dice similarity coefficient (DSC) was calculated for quantitative evaluation.
Results: The average DSCs across four organs were 84.3±6.2%, 89.8±5.9%, 88.7±7.7%, 85.4±11.7%, and 91.1±6.7% for DIR, SAM, MedSAM, TS, and SAM2, respectively. Segmentations without multi-directional exploration resulted in average DSCs of 88.0±6.5%, 85.7±8.3%, and 85.5±12.6% for SAM, MedSAM, and SAM2, respectively.
Conclusion: This study successfully developed a fully automated zero-shot framework for organ segmentation using DIR and multi-directional exploration. Our framework outperformed other implementations of previous foundation models (SAM, MedSAM) and a publicly available model (TS).