3206 - Fine-Tuning SAM for CBCT Segmentation in Offline Adaptive Cervical Cancer Radiotherapy with Limited Data

Purpose/Objective(s):

In definitive radiotherapy for cervical cancer, bladder filling significantly affects the mobility of the uterus and cervical mass, thereby impacting interfractional treatment precision. However, efficient offline adaptive radiotherapy (ART) remains challenging due to poor CBCT image quality and anatomical variability on conventional linear accelerators with kilovoltage imaging systems. Here, we investigate whether fine-tuning the Segment Anything Model (SAM) with a custom CBCT preprocessing strategy, can achieve state-of-the-art segmentation accuracy using a small patient cohort, thereby facilitating efficient offline ART workflows.

Materials/Methods: We retrospectively analyzed 17 cervical cancer patients (136 CBCT scans): 12 for model development and 5 for testing, with uterus, cervical mass and bladder as the main target areas (patient-level split). A two-step innovation was implemented: (1) a custom CBCT preprocessing algorithm to enhance soft-tissue contrast and suppress artifacts via adaptive filtering, windowing, and intensity normalization; (2) fine-tuning SAM’s Transformer-based decoder to the medical imaging domain, using only the limited CBCT dataset. Performance was compared to original SAM and nnUNet using Dice (DSC), 95% Hausdorff distance (HD95), average surface distance (ASD), and relative volume difference (RVD). Statistical comparisons were performed using Wilcoxon signed-rank tests (a=0.05).

Results: The proposed model significantly improved DSC, HD95, ASD, and RVD versus nnUNet for both organs (p=0.031). Relative gains over nnUNet were 7.5% (uterus DSC) and 5.8% (bladder DSC), with substantially lower surface errors (e.g., 43% lower ASD for bladder). Segmentation improvements indicate more geometrically precise and clinically reliable contours (Table 1).

Conclusion: This study demonstrates that fine-tuning a foundation model like SAM, augmented by a dedicated CBCT preprocessing pipeline, overcomes the small-data hurdle and achieves high-precision CBCT segmentation for quick evaluation and preparation during offline ART, using limited patient-level training data. By delivering accurate and robust contours, this approach has the potential to reduce manual editing time, minimize inter-observer variability, and streamline offline ART workflows for cervical cancer patients, without upgrading high-value equipment. The paradigm of fine-tuning large pre-trained models opens a new pathway for developing clinically deployable AI tools in data-scarce scenarios.

Table 1. Segmentation performance (mean ± SD).

Organ	Metric	Original SAM	nnUNet	Proposed (fine-tuned SAM)
Uterus	DSC	0.69 ± 0.06	0.80 ± 0.03	0.86 ± 0.03
	HD95 (mm)	11.32 ± 2.47	6.60 ± 0.81	5.39 ± 1.38
	ASD (mm)	6.19 ± 1.56	3.46 ± 0.46	2.32 ± 0.60
	RVD (%)	93.59 ± 29.45	45.52 ± 7.40	10.01 ± 16.36
Bladder	DSC	0.81 ± 0.05	0.86 ± 0.04	0.91 ± 0.03
	HD95 (mm)	18.67 ± 4.95	14.41 ± 4.38	8.62 ± 2.80
	ASD (mm)	6.99 ± 1.69	4.86 ± 1.09	3.03 ± 0.72
	RVD (%)	35.89 ± 14.49	12.35 ± 8.36	8.11 ± 7.29