Segmentation-Guided CycleGAN for Synthetic CT Generation from MRI for Radiotherapy

MRI and CT images are fundamental diagnostic and radiation therapy modalities. MRI is characterized by high-quality soft tissue contrast. However, it lacks the electron density information required for dose calculations. On the other hand, CT is considered the best choice in radiation therapy planning thanks to its accurate electron density data. Still, it exposes patients to the radiation. Recently, generating sCT from MRI using deep learning has acquired significant attention as an effective solution that mitigates the risks of using multiple imaging modalities during the treatment simulation phase and shows considerable benefits. The current deep Learning-based methods suffer from blurred outputs when MRI-CT misalignments occur. To overcome these limitations, this work proposes a framework employing a universal segmentation model (USM) to align the CycleGAN-based model. To align the generated CT images, we calculated differences between the target and the generated image segmentation maps using USM. This alignment ensures a full representation of the complexity of medical scans and enables the model to capture more information. Our results show that our method outperforms all baselines and improves the synthesis accuracy by a sianificant margin in sCT generation.