197 - SAM-ART: An Efficient Segmentation Method with a Large Vision Model for MRI-Guided Adaptive Radiotherapy

B. Yang, Y. Liu, N. Lu, K. Men, and J. Dai; Department of Radiation Oncology, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China

Purpose/Objective(s): Autosegmentation methods such as deformable image registration (DIR) and traditional deep learning (tDL) in MRI-guided adaptive radiotherapy (MRIgART) often require significant manual editing. Although the Segmentation Anything Model (SAM) exhibits impressive capabilities as a large vision model (LVM), its application in medical images necessitates numerous human interactions and yields suboptimal results. To address this, we introduce SAM-ART, an efficient segmentation method tailored for MRIgART tasks. Materials/

Methods: 38 patients with prostate cancer participated in this study, providing planning CT (pCT) scans, corresponding approved contours, and daily MRI (dMRI) scans. SAM-ART comprises an image encoder, a prompt encoder, and a mask decoder. Unlike SAM, SAM-ART incorporates an adapter module within the image encoder for domain adaptation to medical images. Using mask and box prompts, SAM-ART propagates contours from pCT to dMRI using DIR and employs them as mask prompts, offering patient-specific information. To mitigate the inaccuracies inherent in DIR, box prompts are used in slices prone to false positive (FP) or false negative (FN) predictions. During training, the parameters of the image encoder remain frozen, with fine-tuning applied solely to the adapter module, prompt encoder, and mask decoder. A 5-fold cross-validation was conducted, comparing SAM-ART with DIR, tDL, tDL + mask, and SAM-ART using different prompts (point or box). The performance was evaluated on segmentation of the clinical target volume (CTV).
Results: The results indicate enhanced segmentation accuracy with proposed SAM-ART, exhibiting a mean Dice value of 93.29%, surpassing DIR (87.41%, p < 0.001) and tDL (80.70%, p < 0.001). Comparing different prompts in SAM-ART, the proposed mask + box prompts (93.29%) outperformed the other modes (point: 92.27%, p < 0.001, box: 92.78%, p < 0.001. Sole reliance on box or point prompts necessitated manual interactions on each slice, whereas using only mask prompts introduced FP and FN predictions. Mask + box prompts effectively mitigated FP/FN predictions with minimal manual intervention.
Conclusion: This study introduces a novel segmentation method that leverages LVM and integrates patient-specific information through mask prompts. Outperforming baseline methods, SAM-ART enhances the efficiency of ART. Abstract 197 – Table 1: Performance of different methods (HD: Hausdorff Distance, MDA: mean distance to agreement)

Methods		Dice (%)	HD (mm)	MDA (mm)
DIR		87.41 ± 3.93	4.11 ± 0.94	0.39 ± 0.25
tDL		80.70 ± 6.31	6.19 ± 1.36	0.63 ± 0.43
SAM-ART	point	92.27 ± 1.54	2.93 ± 0.56	0.13 ± 0.07
	box	92.78 ± 1.39	2.79 ± 0.44	0.11 ± 0.06
	mask + box	93.29 ± 1.15	2.60 ± 0.39	0.09 ± 0.04