A Surfaced-based Predictive Model with Interpretability on Cortical Features
Poster No:
1900
Submission Type:
Abstract Submission
Authors:
Zhuoshuo Li1, Youbing Zeng1, Jiaying Lin1, Duan Xu2, Hosung Kim3, Mengting Liu1
Institutions:
1Sun Yat-sen University, Shenzhen, Guangdong, 2University of California San Francisco, San Francisco, CA, 3University of Southern California, Los Angeles, CA
First Author:
Co-Author(s):
Introduction:
Utilizing machine learning models to predict phenotypes, such as age, gender, and brain disease states, based on the highly folded cerebral cortical surface and its morphological features, stands as a crucial method in surface-based analysis and its practical applications. It significantly contributes to investigating brain development/aging trajectories and developments of brain disorders [1]. In the current research, surface-based models exist but have limitations in flexibly identifying the best sub-graph structure or important vertices that contribute best to the prediction task and conducting the interpretable results at the feature level. These limitations are notable because morphological features correspond to numerous phenotypes and functions, and different regions exhibit diverse responses to various predictive demands [2]. To address these, in this study, we propose a surface-based predictive model that focuses on a self-interpretable feature, namely the Surface Graph Neural Network (surfGNN).
Methods:
We consider the surface mesh with individual cortical feature as a sparse graph, and the quantity of sparse graphs for each subject is consistent with the number of cortical features extracted for a given task. Based on this, the surfGNN model is constructed, as Fig. 1 shows. The entire framework of surfGNN consists of context-aware learning structures and a regional-specific learning structure for each cortical feature, and a score-weighted fusion structure across all features for prediction.
The context-aware learning structure is utilized to condense the topology structure of a low-level surface mesh (with a larger number of vertices). The regional-specific learning structure is employed on a high-level surface mesh (with a lower number of vertices), to accurately identify and emphasize important contributions from various regions. Moreover, a novel score-weighted fusion mechanism is proposed to amalgamate node information derived from individual cortical features within the graph learning framework. This mechanism also facilitates the creation of node scores, providing interpretable results that are specific to each cortical feature.
The context-aware learning structure is utilized to condense the topology structure of a low-level surface mesh (with a larger number of vertices). The regional-specific learning structure is employed on a high-level surface mesh (with a lower number of vertices), to accurately identify and emphasize important contributions from various regions. Moreover, a novel score-weighted fusion mechanism is proposed to amalgamate node information derived from individual cortical features within the graph learning framework. This mechanism also facilitates the creation of node scores, providing interpretable results that are specific to each cortical feature.
·Fig.1. Overview of the proposed network architecture and several key modules.
Results:
We apply the surfGNN to a neonatal brain age prediction task on a dataset with morphological features including cortical thickness, sulcal depth, and gray matter/white matter (GM/WM) intensity ratio. The dataset consists of harmonized MRI images of 481 subjects (503 scans) from UCSF and dHCP [3] cohorts, where the cortical surface is reconstructed using the NEOCIVET pipeline [4] and downsampled to obtain sparse maps at multi resolutions with distinct quantities of vertices. Our model achieves the best MAE of 0.827±0.562 weeks of post-menstrual age. In Fig. 2(a), we assess our model's predictive performance against state-of-the-art surface-based models. Our surfGNN model demonstrates superior predictive accuracy with sparse graph inputs at each resolution. In Fig.2(b), it provides the activation maps illustrating variations in the response of distinct brain regions to predicted outcomes across various cortical features.
·Fig.2. (a) Performance comparison with different models on different resolution of the input sparse graphs. (b) The spatial activation maps for the three cortical features.
Conclusions:
We have presented a novel surface-based predictive model with self-interpretability, which could fully integrate both the topological structure traits and the physiological characteristics of the morphological features on the cortical surface. Our model showcases exceptional predictive performance and generates effective activation maps at a feature level. The limitations of this study encompass the limited scale of the experimental datasets, the lack of tests in diseased cohorts, and the finite precision of the interpretable results. Future work will focus on addressing these issues.
Lifespan Development:
Early life, Adolescence, Aging
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Methods Development 1
Keywords:
Cortical Layers
Machine Learning
Morphometrics
STRUCTURAL MRI
1|2Indicates the priority used for review
Provide references using author date format
[2] J. H. Gilmore, et al. 'Imaging structural and functional brain development in early childhood.', Nature Reviews Neuroscience 19.3 (2018): 123-137.
[3] A. D. Edwards, et al. 'The developing human connectome project neonatal data release.' Frontiers in Neuroscience 16 (2022): 886772.
[4] M. Liu, et al. "Robust cortical thickness morphometry of neonatal brain and systematic evaluation using multi-site MRI datasets." Frontiers in Neuroscience 15 (2021): 650082.