Author: Yang Lei, Tian Liu, Ren-Dih Sheu, Meysam Tavakoli, Jing Wang, Kaida Yang, Jiahan Zhang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, Department of Radiation Oncology, Emory University 🌍
Purpose:
The study aimed to improve target and organ at risk (OAR) segmentation in low-dose-rate brachytherapy (LDR-BT) for prostate cancer treatment, by integrating clinical guidelines into deep learning workflow via task-specific prompt generation using large language models (LLMs). This method addresses the critical need for precise target and OAR delineation, which is essential for optimizing treatment planning and minimizing radiation-induced toxicity.
Methods:
A Swin transformer-based encoder was used to extract features from planning CT images. American Brachytherapy Society (ABS) guideline-derived prompts, generated by LLMs and encoded through a text encoder, provide content-sensitive anatomical and clinical guidance. This approach directed a fully convolutional network-based encoder to aid precise multi-task OAR segmentation. The Contrastive Language-Image Pretraining (CLIP) neural network was used to correlate text with images, generating vector representations to guide the segmentation tasks dynamically. The model's performance was evaluated using data from 20 prostate cancer patients who previously received LDR-BT. The Dice similarity coefficient (DSC), measuring segmentation overlap, and other metrics were analyzed using paired t-tests to assess statistical significance.
Results:
The prompted segmentation demonstrated significant improvements across all evaluated pelvic structures. DSC of the prostate increased from 0.862±0.032 without prompts to 0.889±0.024 with prompts, a 3.21% improvement (p<0.0001). Bladder DSC improved by 2.22% (0.859±0.203 to 0.878±0.195, p=0.0103). The Urethra exhibited a 23.42% DSC improvement (0.48±0.148 to 0.553±0.130, p< 0.0001). Rectum DSC improved by 11.45% (0.738±0.077 to 0.823±0.078, p<0.0001). Most notably, the seminal vesicles showed a 333.47% improvement in DSC (0.183±0.133 to 0.792±0.074, p<0.0001). Other metrics (HD95, MSD, CMD, and VD) also demonstrated consistent improvements, particularly for highly irregular structures like the seminal vesicles.
Conclusion:
The integration of clinical guidelines through LLM-generated prompts significantly enhances male pelvic organ segmentation accuracy. This method can be applied to other anatomical sites to improve and refine auto segmentation results.