Author: Eric Chang, Nguyen Phuong Dang, Andrew Lim, Lauren Lukas, Lijun Ma, Yutaka Natsuaki, Zhengzheng Xu, Hualin Zhang 👨🔬
Affiliation: Radiation Oncology, Keck School of Medicine of USC 🌍
Purpose: Harnessed the power of AI and Deep Learning (DL), Generalized Neural Network models for medical image transformation are trained to predict target images from reference images, often requiring paired training data of any modalities on the same patient. However, the availability of such paired datasets in clinical settings is a limiting factor in further developing clinical applications. This study addresses two key questions: 1) Are these transformation models reversible? 2) Can they be stacked in series? By validating these properties, we aim to demonstrate that these models are modular, allowing target image transformations in any direction without the need for paired training data.
Methods: Data from the SynthRAD 2023 Synthetic CT challenge was utilized, consisting of four pre-processed image pair sets (n=180 per set): Brain MR-CT, Pelvis MR-CT, Brain CBCT-CT, and Pelvis CBCT-CT. Three transformation models were developed for each imaging pair (total of six models): MR to Synthetic CT, CT to Synthetic MR, and CBCT to Synthetic CT. These models were built using TensorFlow U-Net on Python, with DL tasks conducted on an Ubuntu Linux system utilizing four Nvidia GPUs (10GB each). The MR-CT models were trained for 350 epochs, and the CBCT-CT models for 100 epochs, with a batch size of 12 images and a minimum of three models per DL task.
Results: The forward and reverse CT-MR models successfully predicted high-quality images for both brain and pelvis datasets. The CBCT-CT model effectively reduced CBCT artifacts, and the subsequent CT-MR model converted the synthetic CT images into comparable synthetic MR images, demonstrating successful image transformations.
Conclusion: This study confirms that medical imaging transformation models are both reversible and stackable in series. These models can be constructed without requiring direct reference-target image pairs, providing enhanced flexibility and applicability for various imaging modalities in clinical practice.