Physics-trained data scientist interested in machine learning, neuroscience, and blockchains

University College London, University of London

Neuroimaging PhD '21, University College London, University of London

Copenhagen University

radiology, nuclear medicine and imaging

human-computer interaction

artificial intelligence

statistics, probability and uncertainty

cognitive neuroscience

pediatrics, perinatology and child health

hardware and architecture

computational theory and mathematics

information systems and management

global and planetary change

preregistration

Leevi Kerkelä's Notebook

Replication of "Training data distribution significantly impacts the estimation of tissue microstructure with machine learning"

1. TitleReplication of "Training data distribution significantly impacts the estimation of tissue microstructure with machine learning"2. AuthorsLeevi Kerkelä3. DescriptionI am proposing to openly attempt to replicate the results of the following study by Gyori et al.: "Training data distribution significantly impacts the estimation of tissue microstructure with machine learning". The original study investigated the impact of training data distributions on the accuracy and precision of microstructural parameter estimation from diffusion magnetic resonance imaging data (dMRI). Despite being published only two years ago, the paper is quite impactful, having been cited over 40 times, according to Google Scholar. The authors seem to have done an excellent job in their analysis, and I have no reason to doubt their results. Instead, I intend to replicate the main results in the true spirit of science. I will share the results on my personal GitHub and ResearchHub.4. SignificancedMRI is an increasingly important imaging methodology that enables the quantification of brain tissue microstructure in vivo. dMRI is regularly used in clinical settings and neuroscience. Analysis of dMRI is challenging, and as a consequence of the recent advances in machine learning (ML), many scientists are investigating if ML can improve the analysis of dMRI data. The work by Gyori et al. highlights that despite being able to produce stunning maps nearly instantly, the performance of ML models depends strongly on the training data distribution. Unfortunately, the authors of the original papers did not share the code or the data, thus making it harder for other researchers or students to investigate further. By openly replicating the study's main results using open-source software and openly available data, I will lower the barrier of entry for contributing to the research of ML applications in dMRI and medical imaging in general. Gyori et al. focused on only one microstructural model, and by replicating it openly, my work proposed here will allow other researchers to investigate the effects of training data distribution on the models they are using.5. HypothesisDoes training data significantly impact the estimation of tissue microstructure with machine learning?6. Study PlanThe study involves the following steps:Set up a public GitHub repository and define the used Python packages (matplotlib, nibabel, pytorch, scipy, etc.) in a reproducible Conda environmentDownload openly available two-shell high-angular resolution diffusion imaging data (e.g., HCP)Preprocess the dataEstimate the microstructural parameters using conventional fittingTrain an artificial neural network to estimate the model parameters using two different training data distributionsCompare the predictions between the two models trained with different training dataAsk an expert and ResearchHub users to review and run the codeWrite up a detailed report in the style of a short academic paper or a conference abstractThe details are explained in the original paper. The work will be done using Python's scientific stack.7. CostsI will use openly available data and free open-source software. The ML models will be trained on the free GPUs of Google Colab. I'm a PhD-level expert in the field, and I am willing to contribute up to 10 hours per week of my free time to work on this project in the evenings, so support from the ResearchHub community would be greatly appreciated. I expect to be able to finish the primary goal of the project in 40 hours.  