Modeling task-based fMRI data via deep belief network with neural architecture search. (July 2020)

• Record Type: Journal Article
• Authors: Qiang, Ning; Dong, Qinglin; Zhang, Wei; Ge, Bao; Ge, Fangfei; Liang, Hongtao; Sun, Yifei; Gao, Jie; Liu, Tianming
• Is Part Of: Computerized medical imaging and graphics. Volume 83(2020)
• Publication Date: 2020-07
• Subjects: Task fMRI -- Neural architecture search -- Deep belief network -- Deep learning -- Unsupervised learning; Diagnostic imaging -- Periodicals; Imaging systems in medicine -- Periodicals; Diagnosis, Radioscopic -- Data processing -- Periodicals; Diagnostic Imaging -- Periodicals; Imagerie pour le diagnostic -- Périodiques; Diagnostic imaging; Periodicals; Electronic journals; Electronic journals; 616.0754
• DOI: 10.1016/j.compmedimag.2020.101747
• Languages: English
• ISSNs: 0895-6111

Abstract

Graphical abstract: Illustration of the proposed NAS-DBN framework for deriving functional brain networks and temporal responses from volumetric task-based fMRI data. (A) PSO based NAS framework. 30 particles are initialized with random position and velocity, each particle represents a sub-net of DBN with specific network architecture. Testing loss of each sub-net (fitness function of PSO algorithm) is calculated after training. Then, evaluation and updating operation is conducted according to standard PSO procedure. After iterations of PSO, the global best particle represents the optimal architecture of DBN. (B) The optimal architecture of DBN. After NAS, DBN with optimal architecture is trained on group-wise fMRI data, and hierarchical temporal responses can be generated from each hidden layer. (C) Functional brain networks (spatial maps). The weight matrix of DBN model are mapped back to the brain 3D space and visualized as FBNs, which will be further compared with GLM derived templates by overlap rate. Highlights: A novel unsupervised PSO based deep belief network model with neural architecture search (NAS-DBN) in modeling functional brain networks (FBNs) from fMRI. The NAS-DBN acts as a hierarchical feature extractor that decompose the preprocessed fMRI data into spatial features (FBNs) and temporal features. The data-driven deep learning model provides much more latent information of fMRI than traditional models. We derived 260 of hierarchical FBNs and temporal features from each hidden layer of DBN, including meaningful task specific FBNs and 10 resting state networks (RSNs). Abstract: It has been shown that deep neural networks are powerful and flexible models that can be applied on fMRI data with superb representation ability over traditional methods. However, a challenge of neural network architecture design has also attracted attention: due to the high dimension of fMRI volume images, the manual process of network model design is very time-consuming and not optimal. To tackle this problem, we proposed an unsupervised neural architecture search (NAS) framework on a deep belief network (DBN) that models volumetric fMRI data, named NAS-DBN. The NAS-DBN framework is based on Particle Swarm Optimization (PSO) where the swarms of neural architectures can evolve and converge to a feasible optimal solution. The experiments showed that the proposed NAS-DBN framework can quickly find a robust architecture of DBN, yielding a hierarchy organization of functional brain networks (FBNs) and temporal responses. Compared with 3 manually designed DBNs, the proposed NAS-DBN has the lowest testing loss of 0.0197, suggesting an overall performance improvement of up to 47.9 %. For each task, the NAS-DBN identified 260 FBNs, including task-specific FBNs and resting state networks (RSN), which have high overlap rates to general linear model (GLM) derived templates and independent component analysis (ICA) derived RSN templates. The average overlap rate of NAS-DBN to GLM on 20 task-specific FBNs is as high as 0.536, indicating a performance improvement of up to 63.9 % in respect of network modeling. Besides, we showed that the NAS-DBN can simultaneously generate temporal responses that resemble the task designs very well, and it was observed that widespread overlaps between FBNs from different layers of NAS-DBN model form a hierarchical organization of FBNs. Our NAS-DBN framework contributes an effective, unsupervised NAS method for modeling volumetric task fMRI data.