Introduction: Magnetic Resonance Imaging is a commonly used technique for the initial diagnosis of gliomas. T1, T2, T2-FLAIR, and post-contrast T1 with gadolinium-based contrast agents (GBCAs) can show tumor characteristics. However, using this contrast agent poses a risk to patients with kidney failures, has environmental impact, and increases cost. To address these issues, we aimed to evaluate the potential of deep learning in generating synthetic post-contrast T1 images without using contrast agents. Method: The project investigates the potential of using deep learning (DL) to generate synthetic post-contrast T1 images based on T1, T2, and T2-FLAIR provided by the Erasmus Glioma Database. Exploring different model architectures, loss functions, and input sizes to discover the optimal approach. Results: Results show that individual loss functions, input size, and model complexity slightly impact the accuracy of synthetic post-contrast T1 images. Combining loss functions, however, was the most promising approach for image generation. Models trained with ℒm could generate low detail enhancement. Resulting in 0.0478±0.0076, 0.0139±0.0036, and 0.879±0.024 for MAE, MSE, and SSIM, respectively. Conclusion: The study's findings indicate that DL is promising for generating synthetic post-contrast T1 images without using GBCAs. However, further research is required to generate realistic synthetic post-contrast T1 images. The study, however, provides a basis for future work and highlights the importance of reducing the use of GBCAs in clinical practice.

Introduction Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI) is a quantitative MRI method, where the diffusion of water molecules is exploited as a source of contrast among tissues. By sensitising DW-MR-measurements to diffusion, we enable the non-invasive quantification of tissue microstructure. The signal observed in a voxel arises from contributions of various diffusion processes within different cellular components (i.e. the tissue microstructure). Many methods exist aiming to disentangle these signal contributions and to compute biologically meaningful microstructure indices. A new, innovative method is using Artificial Intelligence (AI) to directly predict microstructure from DW-MR-signal data, instead of relying on signal representations or biophysical models. Objective The objective of this project is the development of a novel AI-based method to estimate microstructural parameters in the liver, specifically cell size and intracellular fraction, given paired examples of DW-MRI signals and microstructural features as seen on histology. We aim to outperform the estimation of microstructure by a well-established biophysical model. The method will be developed on pre-clinical data, and will be tested by deployment on in vivo patient data. Methods Seven algorithms were implemented in Python, of which four voxel-wise Machine Learning (ML) (Linear Regression, Piecewise Linear Regression, K-Nearest Neighbour, Random-Forest Regres sor (RFR)), one patch-wise ML (PW-RFR), one voxel-wise Deep Learning (DL) (Multi-Layer-Perceptron (MLP)) and one patch-wise DL (PW-MLP). The algorithms were trained in a supervised manner, using MRI-derived data co-registered to histology-derived parametric maps of the targets (i.e. the ground truth) of seven mouse livers. The performance of the algorithms based on three input feature types is analysed: (i) the raw DW-MR-data, (ii) signal representations, (iii) parametric maps computed by fitting a multi-exponential biophysical model. Quantitative evaluation metrics (Mean Absolute Error (MAE) and Relative MAE) and visual inspection are used for performance assessment. The most promising pre-trained algorithm from the pre-clinical study is consequently tested on clinical data of three patients. Results Based on quantitive evaluation metrics and visual inspection of the predictions in comparison with the histology-derived ground truth, the PW-RFR was found to be the most promising algorithm. With the PW-MLP exhibiting the second best results, the importance of incorporating spatial information, here via the patch-wise approach, is supported. No large differences were observed in the performance of the best algorithm, the PW-RFR, for the three input feature types. From the clinical study appeared that the PW-RFR is able to at least partly capture notable contrasts in the lesion microstructure, and the predictions are in a realistic range in comparison with histology-derived microstructure values as obtained via biopsy of the lesions. Conclusion The PW-RFR algorithm demonstrates a promising performance in both the pre-clinical and the clinical experiment. This points towards the feasibility of developing tools that learn how to predict histopathological characteristics from DW-MRI on pre-clinical MRI-histology data, and consequently deploy these tools in a clinical setting. However, these results are preliminary, and further validation on larger and more diverse datasets is required.

Quantitative MR obtains images containing meaningful physical and chemical characteristics of the tissues, allowing for comparison between patients. The procurement of this type of MRI through specific correction acquisition sequences tries to eliminate the maximum possible standardization problems that the technique inherently possesses. However, this does not provide perfect results. Deep learning could be used as a tool to finalize this harmonization after acquisition is finished. In this project, a CycleGAN is used to transfer the style of one specific MRI scanner into the images of another specific scanner. The aim is to achieve a better harmonization that eases posterior image analysis and to possibly solve other issues like hardware obsoleteness. Inspired by the literature, which has never applied image style transfer to this type of images, different methodologies are tested. Some have been applied to other MRI modalities, like an extra similarity measure in the loss function. One novel implementation is tested. It consists on an extra discriminator that tries to reinforce the classification of the original and fake/generated images of one scanner as one class, as opposed to the class formed by the original and fake images of the other scanner. Validation is based on visual inspection; histogram comparison; SSIM, NRMSE and correlation measures and CNN classification of the generated images (in a network trained to distinguish the origin of the scanners). Experiments show the inconclusiveness of the possibility to apply the general CycleGAN loss function to a set of images with such visual similarity. A further study on the specific details that a discriminator uses in order to classify images as coming from a given scanner could help design a specific loss which’s optimization generates the desired results.