TransDose: Transformer-based radiotherapy dose prediction from CT images guided by super-pixel-level GCN
Zhengyang Jiao1, Xingchen Peng2, Yan Wang1
1School of Computer Science, Sichuan University, Chengdu, China.
Medical Image Analysis
|July 23, 2023
Summary
TransDose predicts radiotherapy dose distributions using only CT images, eliminating the need for segmentation masks. This novel approach accelerates treatment planning by leveraging graph convolutional networks and Transformers for anatomical feature extraction and inter-slice continuity.
Area of Science:
- Medical Physics
- Radiotherapy
- Artificial Intelligence in Medicine
Background:
- Radiotherapy treatment planning relies on accurate dose distribution maps, traditionally generated through manual, time-consuming expert processes.
- Existing automated methods often require auxiliary inputs like segmentation masks, limiting efficiency and applicability.
- The need for faster, more efficient radiotherapy planning is critical for clinical practice.
Purpose of the Study:
- To develop a novel automated method for radiotherapy dose distribution prediction using only CT images as input.
- To eliminate the dependency on segmentation masks for tumor and organs at risk (OARs).
- To improve the efficiency and potential application of automated radiotherapy planning.
Main Methods:
- Introduced TransDose, a deep learning model that utilizes CT images as the sole input.
- Employed a super-pixel-based graph convolutional network (GCN) to extract category-specific anatomical features without segmentation masks.
- Integrated a Transformer architecture to capture long-range dependencies and inter-slice continuity between CT slices and dose maps.
Main Results:
- TransDose successfully predicts dose distributions using only CT images, demonstrating the feasibility of eliminating segmentation masks.
- The model effectively extracts necessary anatomical information through the GCN and benefits from inter-slice continuity via the Transformer.
- Extensive experiments on two real datasets validated the generalizability and advantages of the proposed TransDose method.
Conclusions:
- TransDose represents a significant advancement as the first network specifically designed for dose prediction solely from CT images.
- The approach enhances efficiency by removing the need for manual segmentation, accelerating radiotherapy treatment planning.
- The model's performance indicates strong potential for broader clinical adoption and improved patient care.
