Author: Mojtaba Behzadipour, Siyong Kim, Mitchell Polizzi, Richard R. Wargo, Lulin Yuan 👨🔬
Affiliation: VCU Health - Department of Radiology, Virginia Commonwealth University 🌍
Purpose:
The purpose of this study is to develop a method for generating missing tissue in CT scans of patients with large body sizes, where the field of view (FOV) of the scanner fails to capture the entire anatomy. Missing tissue can cause inaccurate dose calculations, potentially compromising treatment planning accuracy.
Methods:
CT data from the NSCLC Radiomics dataset (422 patients) was used. The dataset includes pre-treatment CT scans and manual delineations of gross tumor volumes. Python scripts were developed to remove the treatment couch from images, ensuring the focus remained on the patient anatomy. HU values were examined to segment bone and soft tissue into separate window-level datasets. These were saved as 8-bit PNG images. The dataset was split into 380 patients for training and 42 for testing a GAN model which uses a transformer-based architecture with a two-stage process: a coarse network for initial content generation, followed by a refinement network with an attention-aware layer (AAL) for improved consistency and detail.
Results:
The window-level based training approach demonstrated superior performance in reproducing both soft tissue and bony structures compared to single-piece training. The learning rates for bone and soft tissue components exhibited distinct behaviors, indicating the effectiveness of separate processing.
Conclusion:
This study demonstrates the feasibility of using AI-based methods to generate missing tissue in CT scans, addressing a critical limitation in dose calculations for large patients. The window-level segmentation approach significantly improved the accuracy of regenerating bony and soft tissue structures, suggesting a potential clinical impact in enhancing treatment planning reliability for out-of-FOV cases.