Self-supervised techniques for medical image representation learning
Reference persons FABRIZIO LAMBERTI
External reference persons Lia Morra
Research Groups GR-09 - GRAphics and INtelligent Systems - GRAINS
Thesis type RESEARCH THESIS
Description Training deep neural networks require large scale annotated datasets which are often difficult to collect in the medical domain. As a consequence, a plethora of small and very specialized datasets have emerged. The medical domain would extremely benefit from more general purpose models that can be applied to a variety of problems. A possible strategy to overcome this issue is to leverage self-supervised techniques in which the network is pre-trained by using a pretext task. A simple example is rotation: the image is rotated and the network learns to detect the orientation; more generally, a pretext task can be any for which the labels are automatically generated. If the pre-training dataset is sufficiently large and varied, the features learnt should be sufficiently general to be useful for many downstream tasks. The network is then fine-tuned towards the final task (e.g., lesion detection or segmentation). The goal of this thesis, starting from an existing dataset generated in a previous work, is to systematically explore different self-supervised learning techniques, including generative components, and explore their relationship with the diagnostic/classification tasks. Secondly, we will explore ways to make the features learnt generalize better to different imaging modalities. Strong programming and analytical skills are required, as well as a solid background in machine learning/deep learning.
See also http://grains.polito.it/work.php
Deadline 09/02/2023 PROPONI LA TUA CANDIDATURA