Incorporating Uncertainty Estimation into Deep Learning Methods for Medical Image Analysis

The School of EECS is hosting the following PhD Progress review seminar

Incorporating Uncertainty Estimation into Deep Learning Methods for Medical Image Analysis

Speaker: Chaoyi (Joey) Li
Host: Prof Brian Lovell

Abstract: Deep learning has yielded competitive results for various tasks on medical images such as classification and segmentation. Nevertheless, the lack of reliability and trustworthiness of deep learning methods limit their deployment in real clinical scenarios. In recent years, uncertainty estimation has started to receive increasing attention in the deep learning literature, due to its essential role in producing a confidence evaluation along with the prediction of the deep learning model [1]. It provides additional information to the clinician in the decision-making process.

However, there are rare studies to explore how to utilize uncertainty to enable models to achieve better and more robust performance in the field of medical imaging. The aim of this research is to apply uncertainty to support deep learning techniques to perform better and be more robust when applied to medical image analysis. This will be achieved by optimizing the supervised training strategy, reducing the impact of the distribution shift on multi-modality semi-supervised learning, and solving the out-of-distribution problem (e.g. domain generalization) in medical image analysis, which are all utilized by uncertainty estimation. With the successful completion of this research, obtaining calibrated deep learning models according to uncertainty will be a significant step towards deploying deep learning into realistic medical applications.