Exploring contrast generalisation in deep learning-based brain MRI-to-CT synthesis

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Synthetic computed tomography (sCT) is crucial for MRI-based radiotherapy, with deep learning (DL) models showing promise for its generation.
• Challenges arise from varying MRI protocols across centers, leading to poor DL model generalization and low-quality sCT.
• This necessitates methods to improve the robustness and generalizability of DL models for sCT generation.

Purpose of the Study:

• To investigate the efficacy of domain randomization (DR) in enhancing the generalization capabilities of DL models for brain sCT generation.
• To assess the impact of DR on the accuracy and robustness of sCT generated from unseen MRI sequences.

Main Methods:

• A generative adversarial network (GAN) was trained using CT and various MRI sequences (T1, T2, FLAIR) from 95 patients undergoing radiotherapy.
• A baseline model was trained without DR, and its performance on the unseen FLAIR sequence was compared against a DR-enhanced model.
• Image similarity metrics and the accuracy of sCT-based radiotherapy dose plans were evaluated against the ground truth CT.

Main Results:

• The baseline model exhibited the highest mean absolute error (MAE) on the unseen FLAIR sequence (106 ± 20.7 HU).
• The DR model showed improved performance on FLAIR (MAE = 99.0 ± 14.9 HU) compared to the baseline, though still inferior to a model trained with FLAIR data (MAE = 72.6 ± 10.1 HU).
• DR also led to an improved gamma-pass rate compared to the baseline, indicating better dose plan accuracy.

Conclusions:

• Domain randomization effectively improves image similarity and dose accuracy for sCT generation on unseen MRI sequences.
• DR enhances the robustness of DL models, mitigating the need for retraining when encountering new or varied MRI data.
• This approach supports wider clinical adoption of MRI-based radiotherapy by improving the reliability of synthetic CT generation across different imaging protocols.