2271 - A 3D Deep Learning Prompt-Based SmartSAM Model with Zero-Shot Segmentation for Head and Neck Radiotherapy Application

Purpose/Objective(s): The Segment Anything Model (SAM), a recent breakthrough in foundational computer vision models, has gained recognition as a highly adaptable tool for natural image segmentation tasks. Despite its versatility, its application to complex radiotherapy images, particularly for delineating head and neck organs at risk(OARs) and tumors, has seen limited investigation and often requires an overabundant number of input prompts to achieve satisfactory precision. In this research, we present SmartSAM, an innovative web-based platform designed to automate the segmentation of multiple OARs and targets in head and neck radiotherapy by combining foundational models with domain-specific techniques to enhance segmentation accuracy.

Materials/Methods: This collaborative, multi-institutional research utilized two datasets for model development and evaluation. The training dataset included non-contrast simulation CT scans from 200 head and neck cancer patients, sourced from a public dataset and our cancer center. Seventy percent of cases were used for training, and thirty percent for testing. The study aimed to segment 25 OARs and the Gross Tumor Volume (GTV). The SmartSAM model was trained on head and neck datasets and compared to conventional SAM-based 2D segmentation. Accuracy was assessed using Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), with comparisons to prior studies. A preliminary web app, built with Python, JavaScript, and HTML, was developed for zero-shot use by physicians and medical physicists.

Results: The results reveal that the 3D SmartSAM Transformer model consistently surpassed the 2D SAM models in most evaluation metrics, achieving the highest Dice and IoU scores, which reflect its superior segmentation precision. For example, on the public dataset, the 3D SmartSAM model attained a median DSC score of 85.36%, significantly outperforming the 2D SAM model, which scored 65.23%. The 3D SmartSAM model demonstrated robust boundary delineation capabilities. Notably, the model achieved a delineation time of less than 3.5 seconds per structure, offering a significant time-saving benefit for clinicians. To evaluate its generalizability, the model was tested on two separate datasets from different institutions and countries, confirming its ability to perform effectively beyond the scope of the training data.

Conclusion: By simplifying the segmentation process and reducing reliance on manual contouring, SmartSAM enhances the efficiency of radiotherapy workflows while ensuring higher accuracy in treatment planning. Its user-friendly web-based platform makes advanced segmentation tools more accessible, enabling physicians and medical physicists to provide more precise and tailored cancer care. This advancement holds the potential to revolutionize clinical practices, improving patient outcomes and broadening the use of automated segmentation in the field of oncology.