Min-Cut Max-Flow for Network Abnormality Detection: Application to Preterm Birth

• Published in: Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Graphs in Biomedical Image Analysis pp. 164 - 173
• Main Authors: Irzan, Hassna, Fidon, Lucas, Vercauteren, Tom, Ourselin, Sebastien, Marlow, Neil, Melbourne, Andrew
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing 2020
• Series: Lecture Notes in Computer Science
• ISBN: 3030603644; 9783030603649
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-60365-6_16

Neuroimaging studies of structural connectomes typically average the data from many subjects and analyse the average properties of the resulting network. We propose a new framework for individual brain-network structural abnormality detection. The framework uses a graph-based anomaly detection algorithm that allows to detect abnormal structural connectivity on a subject level. The proposed method is generic and can be adapted for a broad range of network abnormality detection problems. In this study, we apply our method to investigate the integrity of white matter tracts of 19-year-old extremely preterm born individuals. We show the feasibility to cast the network abnormality detection problem into a min-cut max-flow problem, and identify consistent abnormal white matter tracts in extremely preterm subjects, including a common network involving the bilateral thalamus and frontal gyri.