Learning-Guided Infinite Network Atlas Selection for Predicting Longitudinal Brain Network Evolution from a Single Observation

Ezzine B. E., Rekık I.

10th International Workshop on Machine Learning in Medical Imaging (MLMI) / 22nd International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), Shenzhen, China, 13 - 17 October 2019, vol.11765, pp.796-805 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 11765
  • Doi Number: 10.1007/978-3-030-32245-8_88
  • City: Shenzhen
  • Country: China
  • Page Numbers: pp.796-805
  • Istanbul Technical University Affiliated: Yes


Building accurate predictive models to foresee the temporal evolution of diverse medical data representations derived from healthy or disordered brain images will enable a formidable, yet challenging, leap forward in the fields of neuroscience and neuro-disorders. However, such models remain very scarce. Existing landmark works on predicting follow-up medical data from a single observation have a few drawbacks. First, these were developed only for predicting brain shapes or images, while brain network representations remain untapped. Second, the bulk of such models lies in the selection of reliable atlases in the baseline domain, which act as proxies for the follow-up domains where the missing data live. However, current atlas selection strategies for prediction suffer from two major limitations: (i) they are selected based on their proximity to the testing sample using a pre-defined distance, which might not be robust to outliers and constrains the locality of the high-dimensional data to a fixed bandwidth, and (ii) atlases are selected independently of one another, which overlooks how the importance of an individual atlas is influenced by all the other atlases in the set. To address these limitations, we propose LINAs, the first framework for predicting brain network evolution from a single timepoint using learning-guided infinite network atlas selection in two steps. First, we learn how to select the best atlases in an unsupervised manner by learning an adjacency graph which encodes the pairwise similarities between all atlases. The relevance score of an atlas is estimated using all possible infinite paths connecting it to other atlases in the set, quantifying its representativeness and centrality. Second, we propose to individualize the atlas score to the testing sample by a supervised re-weighting strategy. Our comprehensive experiments on healthy and disordered brain networks demonstrate the outperformance of LINAs in comparison with its variants as well as state-of-the-art methods. LINAs presents the first step towards building connectome evolution models that can be leveraged for developing precision medicine.