Lieu : Université de Nice principalement Université Laval, Québec, Canada · Contrat : CDD · Rémunération : A négocier
PhD position at Université de Nice and Université Laval (Quebec): Deep learning approaches to enhance cardiology analytics
We are seeking a candidate to fill a position in bioinformatics to develop new deep learning approaches dedicated to cardiovascular diseases. These diseases are a major and rapidly growing health care and socio-economic burden and the cost for health care is estimated to be up to 10 billion $US per year for North America alone. Thus, Cardiovascular diseases have a major impact on patients, society and healthcare provision across the world and it urgently needs major advances in screening, diagnosis, risk-stratification, and treatment.
The thesis project will involve the development of new approaches to integrate several types of complex data usually obtained in cardiology, including echocardiography, CT, PET-CT, that may be composed of images and videos, and multi-omics data (transcriptomics and whole exome sequencing data) along with clinical information. Considering the high complexity to determine the disease phenotypes, complex methods will be considered, such as unsupervised deep learning analysis integrating all these data types in order to establish clusters of disease phenotypes. The candidate will implement clustering using deep generative models (e.g. VAE, GAN, or normalizing flow) able to integrate data from echocardiograms, echocardiography, CT/ PET-CT, and multi-omics (gene expression and extracted variants) which supports images of temporal trends (videos).
Once the clusters are defined, some will be manually associated with disease phenotypes by active Learning/budget Learning, depending on the targeted diseases. Then, using input samples composed of only CT/ PET-CT, echocardiography and clinical data to be practical in real world situations and receive a fast diagnosis, two strategies will be explored: (i) Semi-supervised learning, to let the algorithms learn using only manually labelled samples, or (ii), Uncertain Label Classification, by extrapolating the categories of all samples in each cluster using the manually labelled samples. In (i), the candidate will build robust predictive models by performing semi-supervised multi-label classification. In (ii), the candidate will implement a multi-task network based on Localization U-Net or Deep Residual Recurrent Neural Networks. Finally, to enhance the operability of the models and promote their usage, the candidate will implement visualisation/interpretation methods (e.g. grad-CAM, LIME, SHAP…) to provide explanation of the deep learning models decisions to the clinicians. The deliverable will be an interface to show the parts of images or videos of the echocardiogram and CT with colored gradient overlay layers. The coloration density will be based on the participation of the decision of the models.
Create new deep learning approaches to analyse cardiology data
Elaborate methods providing models interpretation
Master degree in computer science or bioinformatics
Machine learning and Deep learning knowledge