Incorporating physicians' contouring style into auto-segmentation of clinical target volume for post-operative prostate cancer radiotherapy using a language encoder.

INTRODUCTION

Post-operative radiotherapy for prostate cancer requires precise contouring of the clinical target volume (CTV) to account for microscopic disease that is invisible in the image. However, the absence of the prostate gland and unclear anatomical landmarks make this process challenging. In addition, the contouring style is greatly influenced by the physician's experience and preferences. This complicates automatic segmentation algorithms that rely on consistent data but also typically need large amounts data from many physicians to develop. This study presents a novel auto-segmentation method incorporating individual physicians' contouring styles to enhance accuracy and reduce manual workload.

METHODS

We develop Text-UNet, a deep learning-based model integrating a text encoder to encode physician-specific contouring styles into a latent vector. This information is combined with CT image features to guide segmentation. A dataset of 824 patients is divided into 699 training, 49 validation, and 76 testing data. The test set includes recent cases from four physicians, while the training set represents data from seven physicians.

RESULTS

Text-UNet achieves an average Dice score of 85.1%, outperforming the baseline UNet (82.0%) and state-of-the-art models. Incorporating physician-specific information improves segmentation consistency and reduces manual contouring effort, demonstrating its potential to enhance the auto-segmentation of post-operative radiotherapy for prostate cancer.

CONCLUSION

By integrating physicians' contouring styles, Text-UNet addresses inter-physician variability, improving CTV auto-segmentation precision. This approach personalizes radiotherapy planning according to physicians and improves treatment efficiency.