GUI-Based Framework for Automated Deep Learning Synthetic CT Generation and Contouring for Canine Adaptive Radiotherapy on a Dual-Robot Radiotherapy System
Abstract
Purpose
KOALA (Kilovoltage Optimized for AcceLerator Adaptive therapy) is a low-cost adaptive radiotherapy device designed for use in resource-limited settings. KOALA includes cone-beam CT (CBCT) imaging characterized by scatter artifacts and unreliable HU values that limit accurate dose calculation. Generating synthetic CTs (sCT) addresses these limitations by providing planning-quality images from CBCT. In addition, manual contouring of targets and OARs remains time-consuming. We developed a GUI framework that combines sCT generation with automated deep learning-based contouring to support CBCT-based treatment planning for KOALA.
Methods
A GUI was developed using PySide6 to manage the CBCT processing workflow. For sCT generation, we developed a 3D U-Net architecture with attention mechanisms and global residual learning to convert canine CBCT to sCT. For automated contouring, an nnU-Net model was trained to segment the PTV, body contour, and OARs. The GUI supports DICOM loading with orientation correction, synchronized multi-planar visualization with adjustable window/level controls, color-coded contour overlays, contour editing tools, and automated export of DICOM RTSTRUCT files.
Results
The GUI enabled end-to-end canine CBCT processing with minimal user intervention while maintaining visual verification. The sCT model produced planning-quality CT images, mitigating HU inconsistencies inherent to CBCT, and achieved a peak Structural Similarity Index (SSIM) of 88%, indicating preservation of anatomical details for dose calculation. Automated contouring results were visualized and verified within the GUI, achieving 89% accuracy. The framework demonstrated generation of planning-ready datasets compatible with the KOALA treatment pipeline.
Conclusion
This work provides an effective platform integrating CBCT-to-sCT conversion and deep learning auto-contouring into a single GUI workflow for the KOALA system. By combining deep learning methods within an interactive interface, the framework addresses key challenges associated with CBCT-based planning in low-resource environments.