Male pelvic multi-organ segmentation using token-based transformer Vnet.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Accurate segmentation of organs in pelvic CT scans is crucial for effective prostate radiation therapy planning.
• Current segmentation methods can be time-consuming and may lack precision.

Purpose of the Study:

• To develop an automated, accurate segmentation method for the prostate and surrounding organs-at-risk in pelvic CT images.
• To enhance the efficiency and precision of prostate radiation treatment planning.

Main Methods:

• A novel deep learning algorithm combining a U-shaped convolutional neural network (CNN) and a Vision Transformer (ViT) was developed.
• The model incorporates a token-based multi-head self-attention mechanism and knowledge distillation for improved feature learning and dependency capture.
• The network was trained and evaluated on two datasets: one from 94 prostate cancer patients and the public CT-ORG dataset.

Main Results:

• The proposed network achieved high segmentation accuracy, with average Dice scores of 0.91 and 0.93 on the respective datasets.
• Excellent performance was also demonstrated in sensitivity (0.90-0.93), precision (0.92-0.93), and low surface distance errors (HD: 3.78-5.82 mm, MSD: 1.16-1.24 mm).
• The method significantly outperformed other state-of-the-art segmentation techniques.

Conclusions:

• The developed token-based transformer network with knowledge distillation provides accurate and reliable multi-organ segmentation in pelvic CT images.
• This automated approach facilitates the prostate radiation therapy clinical workflow by improving segmentation efficiency and accuracy.
• The findings highlight the potential of advanced deep learning architectures for medical image analysis in oncology.