Evaluation of AI-Based Synthetic CT Generation Methods In Pelvic MRI-Only Radiotherapy
Abstract
Purpose
In MRI-only radiotherapy treatment pathways, synthetic CT (sCT) images can be used for dose calculation, but achieving adequate bony detail in the pelvic region remains challenging. This study aimed to evaluate the performance of multiple AI-based algorithms for sCT generation in the prostate region, focusing on anatomical variations and its potential implications for treatment planning.
Methods
We used publicly available radiotherapy data, including CT, MRI, and CBCT images. We evaluated three AI algorithms: Algorithm A (modified nnU-Net), Algorithm B (2.5D CNN), and Algorithm C (GAN-based Pix2Pix-SwinUNet). Key pelvic organs including bladder, rectum, colon, femoral heads and hip were segmented and compared with reference CT. Performance was analyzed using MAE, PSNR, SSIM, DSC, Sensitivity, HD and MASD.
Results
Algorithm A showed the best performance among the methods, achieving high accuracy for bony structures and bladder (DSC up to about 0.97–1.00). Rectal reconstruction was also acceptable (DSC about 0.87), although its accuracy remained lower than that of more stable organs. In contrast, the colon showed the lowest agreement, which is consistent with natural variations and the presence of gas. Algorithms B and C performed poorly in most organs, especially in complex soft tissues.
Conclusion
Overall, Algorithm A, utilizing a nnU-net system performed better than other methods in generating sCT in the pelvic region and its results were promising for many organs. However, the lower accuracy in the rectum compared to more stable structures suggests that reconstruction of complex soft tissues has not yet reached the ideal clinical level. These errors can lead to uncertainties in dose calculations. Therefore, although AI-based sCT has great potential for MRI-based radiotherapy, but further optimization of these algorithms is necessary for safe clinical implementation.