Related Experiment Video
Updated: Dec 17, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
Daily edge deformation prediction using an unsupervised convolutional neural network model for low dose prior contour
Yingxuan Chen1, Fang-Fang Yin1,2,3, Zhuoran Jiang2
1Medical Physics Graduate Program, Duke University, 2424 Erwin Road Suite 101, Durham, NC 27705, United States of America.
Biomedical Physics & Engineering Express
|June 27, 2020
Summary
This study introduces PCTV-CNN, an unsupervised deep learning method that automates and accelerates edge enhancement for low-dose Cone-Beam CT (CBCT) reconstruction. The new approach significantly reduces processing time while maintaining image quality for adaptive radiotherapy.
Area of Science:
- Medical Imaging
- Radiotherapy
- Deep Learning
Background:
- Previous methods for enhancing edge sharpness in low-dose Cone-Beam CT (CBCT) reconstruction, such as the Projection Contrast-guided Target Volume (PCTV) method, relied on iterative deformable registration.
- This conventional registration process is time-consuming and requires significant user input, limiting its clinical applicability for adaptive radiotherapy.
Purpose of the Study:
- To automate and accelerate PCTV reconstruction by developing an unsupervised Convolutional Neural Network (CNN) model.
- To bypass the need for conventional deformable registration in enhancing edge sharpness for low-dose CBCT.
Main Methods:
- An unsupervised CNN with a U-Net architecture was developed to predict deformation vector fields (DVF) for generating on-board contours necessary for PCTV reconstruction.
- The model takes paired 3D image volumes of prior CT and on-board CBCT as input, predicting DVFs without requiring ground truth data.
- The model was initially trained on brain MRI images and fine-tuned using lung Stereotactic Body Radiation Therapy (SBRT) data, with evaluations conducted on lung SBRT patient data.
Main Results:
- The unsupervised CNN model achieved high accuracy, with cross-correlations between predicted and ground truth edge maps averaging 0.88 in both intra-patient and inter-patient studies.
- PCTV-CNN demonstrated comparable image quality to the traditional PCTV method.
- The PCTV-CNN method significantly reduced registration time from 1-2 minutes to 1.4 seconds, automating the process.
Conclusions:
- It is feasible to employ an unsupervised CNN for predicting daily deformations of on-board edge information for PCTV-based low-dose CBCT reconstruction.
- PCTV-CNN shows significant potential for efficiently enhancing edge sharpness in low-dose CBCT, thereby improving the precision of on-board target localization and facilitating adaptive radiotherapy.
