MSc projects – Neuroimage analysis and machine learning

Project 1: Scan-filter: A deep-learning method for automatic classification of scans with cortical infarcts

Supervisors

Background

In this project, we are interested in accurate classification of brain MRI scans that contain cortical infarcts. Brain infarcts seen on MRI scans are important indicators to the clinical assessment of stroke and neurodegenerative disease, in which each type of infarcts such as cortical and lacunar infarcts may be indicative of disease severity and possible treatment of the disease (see the Fig1 in left). Currently, clinical routine in the assessment of brain infarcts is mostly performed by manual inspection. This has rater variability, is highly dependent on rater experience, and time intensive. To support clinical assessments and subsequent imaging analysis of brain infarcts, fast and automatic methods for identification of the scans containing interested lesions are needed.

Research question and methods

Existing deep learning-based classification methods often output a deterministic class-prediction directly by encoding high-dimensional features into a one- or multi-class decision2 (Fig. 1(a)). In such direct-mapping models, the extracted latent features are often not interpretable beyond visualizing the salience map. Also, we expect the generalizability problem of deep-learning methods to be even typical in such models due to a lack of regularization while having a high degree of freedom. To deal with these problems, we propose a Variational AutoEncoder (VAE)3,4,5-based method for interpretable and robust image classification (Fig. 1(b)). We hypothesize that the compact and regularized feature-representation learned by the VAE model can lead to a more robust prediction on new data than the baseline method (Fig. 1(a)). Feel free to contact us for detailed information of the method and the research plan.

Fig.1 (a) a typical learning-based, and

(b) a VAE-based framework for image classification

Materials

The method will be optimized in a supervised manner using labels of scans, and using a large multi-center dataset, including 3D MRI scans from Heart Brain Study (n=846), Trace-VCI Study (n=832), Rotterdam Study, and some public dataset (n>229).

Reference

[1] Wardlaw et al., The Lancet Neurology 2013. [2] Bron et al., 2020, https://arxiv.org/pdf/2012.08769.pdf. [3] Kingma and Welling, 2013. arXiv preprint arXiv:1312.6114. [4] An example implementation using Keras: https://keras.io/examples/generative/vae/ [5] A blog illustration on VAE methods: https://www.jeremyjordan.me/variational-autoencoders/

Project 2: Vascular bias in Alzheimer’s disease classification

Supervisors

Motivation

About 50 million people worldwide are living with dementia and this number is increasing. Early diagnosis of dementia and its underlying diseases is complex but highly important for delivering the right care to the right patients and for patient selection for clinical trials testing novel treatments. Alzheimer’s disease is the most prevalent cause of dementia, but in most patients coexists with vascular pathology causing cognitive problems.

Machine learning methods classifying Alzheimer’s disease based on imaging data have achieved high performances in a research setting. But performance is much lower in pre-clinical dementia groups (subjective cognitive decline, mild cognitive impairment) and in clinical groups without pre-selection (e.g., for vascular events and risk factors). This might be related to these vascular effects being mostly ignored in many research settings.

Aims

In this project, we aim to study to what extent differences in white matter hyperintensity load between populations hampers generalization for machine learning methods for Alzheimer’s disease classification. In addition, we propose a classification method that takes account of white matter hyperintensities by adding the raw FLAIR scan and/or the processing white hyperintensity map to the classification process.

Project tasks

  • Literature survey on the role of small vessel disease markers, such as white matter hyperintensities, in Alzheimer’s disease classification
  • Quantification of white matter hyperintensities in the ADNI dataset (using existing software).
  • Comparison study of classification performance of convolutional neural network and support vector machine classifiers (Bron et al, 2020) in groups with and without a high white matter hyperintensity load, for diagnosis of Alzheimer’s disease and prediction of Alzheimer’s disease in patients with mild cognitive impairment. (Alternative for diagnosis of cognitive impairment, i.e. To what extent does the MRI scan explain cognitive impairment?)
  • To minimize the vascular bias, adjust the convolution neural network classifier so that it takes additional input from FLAIR scans and/or white matter hyperintensity maps. Validate the updated method in the ADNI data.
  • Option for external validation on the Parelsnoer dataset.
  • Option for clinical supervisor in radiology and/or neurology.

Reference: Bron et al., Cross-Cohort Generalizability of Deep and Conventional Machine Learning for MRI-based Diagnosis and Prediction of Alzheimer’s Disease, 2020, https://arxiv.org/abs/2012.08769