University of Twente Student Theses


Contrast generalisation in deep learning-based brain MRI-to-CT synthesis

Nijskens, Lotte (2022) Contrast generalisation in deep learning-based brain MRI-to-CT synthesis.

[img] PDF
Abstract:Background and purpose: Computed tomography (CT) is the basis for radiotherapy (RT) planning, providing information on electron density needed for dose calculations. Magnetic resonance imaging (MRI) has superior soft tissue contrast and is helpful in tumour delineation, but, unlike CT, it does not inherently contain electron density information. Consequently, CT and MRI are often combined in the planning workflow, requiring image registration. MR-only based RT has been proposed to avoid residual errors after registration, reduce patients' exposure to ionising radiation and simplify the workflow. Synthetic CT (sCT) must be generated to enable MRI-based RT planning, which is possible within seconds using deep learning (DL). MRI acquisition protocols may change over time or differ between centres. Without network re-training, DL models poorly generalise to new domains, including different acquisition protocols, hindering their widespread implementation. Domain randomisation is a learning method that involves generating training data with randomised parameters. The method showed promising results for improving generalisation in, e.g., a segmentation task. This work investigates the ability of DL models for brain sCT synthesis to generalise to MRI scans acquired with unseen sequences without network re-training and how domain randomisation affects model performance on unseen sequences. Materials and methods: Data from 95 patients undergoing RT were included from a retrospective database, requiring a CT image and corresponding T1-weighted MRI with and without contrast (T1wGd and T1w), T2-weighted (T2w) and FLAIR MRI. A Baseline conditional generative adversarial network was trained with and without an unseen sequence (FLAIR) to test how a model performs on the unseen sequence without domain randomisation. Also, two domain randomisation approaches were compared: 1) using synthetic training images with random contrast generated from segmentations of acquired MRI and 2) training on random linear combinations of two MRI sequences. The best approach regarding image similarity between sCT and CT was chosen for comparison with the Baseline models. In a final comparison, image similarity and accuracy of sCT-based dose plans were assessed. Results: The Baseline model trained on T1w(Gd) and T2w images achieved better image similarity on the validation set's FLAIR images than a model trained only on T1w(Gd) images. The domain randomisation method of adding random contrast images resulted in better image similarity on the validation set than the model using linear combination images and was adopted. Of the models included in the final comparison, the Baseline model had the poorest performance on FLAIR, with mean absolute error (MAE) = 106 +/- 20.7 HU (mean +/- SD). Performance on FLAIR significantly improved for the Domain Randomisation model with MAE = 99.0 +/- 14.9 HU. Still, it was inferior to the performance of the Baseline+FLAIR model, trained by adding FLAIR images to the training set (MAE = 72.6 +/- 10.1 HU). The Domain Randomisation and Baseline+FLAIR models resulted in a slight increase in MAE on the seen sequences compared to the Baseline model. The 3D gamma-pass rates were > 95 % for all models and sequences. The 3D gamma-pass rate with 1%,1mm criterion obtained for the Domain Randomisation model for FLAIR images was significantly higher than that obtained for the Baseline model (99.2 +/- 0.9 % vs 99.0 +/- 1.1 %), yet lower than that obtained for the Baseline+FLAIR model (99.4 +/- 0.8 %). Differences in pass rates obtained for the seen sequences between the Domain Randomisation and Baseline model were insignificant. Conclusions: Even without domain randomisation, a satisfactory dosimetric accuracy could be obtained when training on a mix of acquired sequences, even for an unseen sequence. However, domain randomisation improved performance (image similarity and dose accuracy) on the unseen sequence compared to a model trained only on acquired MRI, indicating that the method could help reduce the need for network re-training if the model is to be used on a sequence unseen during network training.
Item Type:Essay (Master)
UMC Utrecht, Utrecht, The Netherlands
Faculty:TNW: Science and Technology
Subject:44 medicine, 50 technical science in general, 54 computer science
Programme:Biomedical Engineering MSc (66226)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page