2259

View Abstract / View Presentation

A semi-supervised graph convolutional network for early prediction of motor impairments in very preterm infants using brain connectome

Hailong Li¹, Ming Chen^1,2, Jinghua Wang³, Nehal A. Parikh^4,5, and Lili He^1,5
¹Imaging Research Center, Department of Radiology, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ²Department of Electrical Engineering and Computer Science, University of Cincinnati, Cincinnati, OH, United States, ³Deep MRI Imaging Inc., Lewes, DE, United States, ⁴The Perinatal Institute and Section of Neonatology, Perinatal and Pulmonary Biology, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ⁵Department of Pediatrics, University of Cincinnati College of Medicine, Cincinnati, OH, United States

Taking advantage of labeled and unlabeled data, a semi-supervised graph convolutional network model is able to identify motor impairments in very preterm infants at 2 years using brain structural connectome obtained at term-equivalent age with an accuracy of 68.1% and an AUC of 0.67.

Figure 2. Overview of semi-supervised graph convolutional network to predict motor impairments in very preterm infants. The input is a cohort graph that describes the inter-subject similarities among training (both labeled and unlabeled) and testing (subjects-to-classify) subjects. The first graph convolutional layer contains 16 graph convolutional filters. The second graph convolutional layer contains 8 graph convolutional filters. The number of graph convolutional filters were selected from empirical values [8, 16, 32, 64] based on validation performance.

Figure 1. (A) Cohort graph describes the inter-subject similarities among the training (both labeled and unlabeled) and testing (subjects to classify) subjects/nodes. (B) Structural connectome feature vectors are used as the node features in the cohort graph. (C) Diffuse white matter abnormality volumes and global brain abnormality scores are used to calculate the inter-subject similarity as edge weights in the cohort graph.