Medical Image Synthesis for Expediting the Diagnosis to Treatment Pathway In Radiation Oncology
Abstract
Purpose
To expedite the diagnosis-to-treatment workflow in radiation oncology, this work evaluates a machine learning approach for generating synthetic radiation therapy (RT) planning images directly from diagnostic computed tomography (CT) images, potentially eliminating redundant planning sessions.
Methods
Paired diagnostic and RT planning CT images for 80 prostate cancer patients were used (split: 70% training, 10% validation, 20% testing). A Cycle Generative Adversarial Network (CycleGAN) with a custom loss function was trained to generate synthetic RT planning images. Image synthesis performance was evaluated using: 1) technical image metrics (i.e. Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Mean Absolute Error (MAE)) to assess voxel-wise similarity to ground-truth planning images, and 2) segmentation-based metrics (i.e. Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD)) to determine the concordance of organ-at-risk segmentations between synthetic images and ground truth planning images using the same automated image segmentation method.
Results
The synthetic RT images achieved a PSNR of 20.21 dB ± 0.61 and SSIM of 0.65 ± 0.03. The synthetic images improved Hounsfield Unit (HU) accuracy compared to standard diagnostic images, reducing the MAE from 76.31 HU to 72.34 HU (mean improvement: 3.97 HU; p < 0.05). Segmentation of the bladder on synthetic images demonstrated high anatomical fidelity to the source diagnostic input (Mean DSC: 0.88 ± 0.10; Mean HD: 7.77 ± 3.62 mm). Comparisons with ground-truth planning images yielded lower agreement (mean DSC: 0.64 ± 0.19; mean HD: 13.39 ± 5.46 mm), expected due to the inter-scan physiological variability in bladder filling.
Conclusion
This work demonstrates the feasibility of using machine learning to generate synthetic RT planning images from diagnostic CT images. Domain shifts are corrected while preserving patient anatomy. Deviations from the ground-truth primarily reflect stochastic physiological changes, which will be explored in future work to improve model robustness.