Unsupervised pattern identification in biomedical imaging data

Background:

Modern imaging techniques can create a large amount of imaging data in a short period of time. However, the vast majority of this imaging data is unlabeled by nature and thus often requires laborious labelling if modern deep learning techniques are to be used to analyse the data. The development of an algorithm with the ability to search through such large biomedical imaging datasets for specific objects/patterns using only a few given examples would address a major problem in computer vision and might have direct impact on multiple applications, such as pathogen identification, quantitative analysis and diagnostic. Even small improvements that reduce the amount of labelled data required would benefit many different applications.

Aim/s:

The aim of the project is to develop an algorithm that identifies structures in biomedical imaging datasets from only very few examples of the data. Two applications are going to be explored: the identification of pathogens in an image data set and the differentiation of different types of cancer. Large data sets are available for both applications.

AI methods:

As no currently existing single un- or semi-supervised method will be sufficient to solve the problem, a combination of multiple deep learning methods is about to be explored, ranging from semi-supervised learning, synthetic data, anomaly detection, vision foundation models, self-supervised learning, human-in-the-loop and others.

Keywords:

Artificial Intelligence, Computer Vision, Biomedical Image Analysis, Deep Learning, Public Health

to the top