Generalizability in multi-centre cardiac image analysis with machine learning

Abstract

[eng] The field of Artificial Intelligence (AI) has undergone a revolution in recent years with the advent of more efficient computing hardware and well-documented software for model development. Many fields are being transformed. Medicine is one of the fields that has seen the appearance of models that can solve complex tasks such as automatic image segmentation or diagnosis. However, there are important challenges that need to be overcome for a successful application in clinical practice. One important challenge is the generalization of models to unseen domains independently of other factors, such as the scanner manufacturer, the scanning protocol, the sample size or the image quality. In this thesis, we aim to investigate the effects of the domain shift in medical imaging, specifically for cardiovascular studies, which present a particular challenge since the heart is a moving organ. Furthermore, we aim to contribute to methods to overcome or reduce the model performance gap.

First, we establish a collaboration with clinical researchers from six different centres from three countries and assemble a large multi-centre dataset to tackle one of the greatest challenges in research: the domain gap problem. We process and annotate the data and develop a benchmark study by organizing an international competition to compare and analyse different techniques to bridge the generalization gap. The dataset is later open-sourced to foster innovation within the research community, becoming the first open multi-centre cardiac dataset.

Then, we perform an exhaustive comparison of domain generalization and adaptation methods, including the best-performing methods in the aforementioned competition, for late gadolinium- enhanced image segmentation for the first time. We show that extensive data augmentation is very important for generalization and that model fine-tuning can reach or even surpass multi-centre models.

Finally, we investigate the effects of differences in image appearance for the first time in a multi-centre study with cardiovascular imaging and compare several harmonisation techniques both at the feature and image levels for improved diagnosis. We show that histogram matching-based harmonisation results in image features (radiomics) that are more generalizable across centres.