Semi-supervised Learning for Medical Image Analysis
Abstract
Semi-supervised learning has recently become a major focus of the deep learning community with recent advancements in methods for self-supervised training methods which can extract useful information from large amounts of data without any manual annotations. These methods are especially useful for the medical imaging community where annotation of a dataset e.g., Computed tomography angiography (CTA) volumes requires specialized radiologists and radiographers to spend multiple hours just to annotate a single image. In this project, we want to utilize the abundant unannotated data for training segmentation models for 3D CTA images. To this end, we are investigating self-supervised training methods for extracting useful features from volumetric data. We are also investigating generation of pseudo-labels for unannotated images using methods like image registration which would be useful for augmenting small annotated medical datasets.
