Semi-supervised Learning for Medical Image Analysis
||Semi-supervised Learning for Medical Image Analysis|
||David Hagerman Olzon <firstname.lastname@example.org>|
||Chalmers tekniska högskola|
||2021-11-11 – 2022-06-01|
Most recent successes of machine learning have been based on Supervised Learning (SL) methods, fueled by large quantities of parallel compute power and humanly annotated training data. However, that option quickly becomes intractable due to the labor-intensive work of manual annotation, especially for medical image data. Instead, many believe that Semi-Supervised Learning (SSL) will drive the next AI revolution by using vast amount of unlabelled data (and some labelled examples) to discover all concepts and underlying causes that matter when interpreting an image. In this project, we will develop new methods and techniques for SSL and apply it to medically relevant problems where lots of image data is available.
The research questions this project will investigate are the following:
1. Can the existing State-of-the-art (SOTA) SSL methods be improved in upon, especially in the aspects of the total data required and the time it takes to converge?
2. Can we supplement self-trained models with additional data from similar medical domains to further improve on fine-tuning tasks?
3. Can we adapt the existing SOTA vision transformer architectures for usage in volumetric data setting?
We will initially focus on lesion detection and semantic segmentation of CT and MRI scans. These are 3D volumes that can be used to examine various organs. Importantly, they can be used to analyze, for instance, possible atherosclerosis in the coronary arteries, which in turn can predict the risk of myocardial infarction in the future.