An explainable deep learning framework for characterizing and interpreting human brain states

Highlights

• Domain knowledge informed SAGPool is proposed to characterize the brain states.

• Dense individualized connectivity-based cortical landmarks are used as graph nodes.

• Dictionary learning is utilized to measure functional interactions among graph nodes.

• Graph nodes which most contributed for the multitask classification can be obtained.

• Learned features can elucidate the most distinguishing brain regions among tasks.

Abstract

Deep learning approaches have been widely adopted in the medical image analysis field. However, a most of existing deep learning approaches focus on achieving promising performances such as classification, detection, and segmentation, and much less effort is devoted to the explanation of the designed models. Similarly, in the brain imaging field, many deep learning approaches have been designed and applied to characterize and predict human brain states. However, these models lack interpretation. In response, we propose a novel domain knowledge informed self-attention graph pooling-based (SAGPool) graph convolutional neural network to study human brain states. Specifically, the dense individualized and common connectivity-based cortical landmarks system (DICCCOL, structural brain connectivity profiles) and holistic atlases of functional networks and interactions system (HAFNI, functional brain connectivity profiles) are integrated with the SAGPool model to better characterize and interpret the brain states. Extensive experiments are designed and carried out on the large-scale human connectome project (HCP) Q1 and S1200 dataset. Promising brain state classification performances are observed (e.g., an average of 93.7% for seven-task classification and 100% for binary classification). In addition, the importance of the brain regions, which contributes most to the accurate classification, is successfully quantified and visualized. A thorough neuroscientific interpretation suggests that these extracted brain regions and their importance calculated from self-attention graph pooling layer offer substantial explainability.

Introduction

Exploring an explainable and generalizable approach for effectively interpreting the brain states is one of the major challenges in the field of brain imaging analysis. In recent years, extensive efforts have been proposed to effectively differentiate and interpret brain states, which contribute to further investigating the working mechanism of the brain (Ganesan et al., 2021). Deep learning approaches, which have been widely utilized in the field of the brain imaging analysis, attract much attention nowadays. For the research field of brain states analysis, deep learning-based algorithms can be roughly summarized into three categories. The first category is the Convolutional Neural Networks (CNN)-based algorithms (Zou et al., 2017; Huang et al., 2018; Zhao et al., 2018; Liu et al., 2019; Zhao et al., 2020). The advantage of adopting the CNN-based algorithm is the ease of modeling and extracting high-dimensional features from the voxel signals of the brain, e.g., Huang et al. designed a CNN-based deep convolutional autoencoder (DCAE) for extracting hierarchical features of brain networks from task fMRI data (Huang et al., 2018). The second category is the Recurrent Neural Network (RNN)-based algorithms (Hjelm et al., 2018; Wang et al., 2019; Cui et al., 2019, 2020), and the advantages of utilizing the RNN-based algorithm is the efficient use of temporal information, e.g., Cui et al. designed deep recurrent neural network to identify brain networks from multiple time scales (Cui et al., 2019). There are also other deep learning algorithms falling into the third category, e.g., Deep Belief Network (DBN) and Restricted Boltzmann machine (RBM) (Dang et al., 2017; Zhang et al., 2019; Dong et al., 2020; Qiang et al., 2020). These methods are highlighted by their attempts to decompose the fMRI signals into components and the corresponding spatial patterns. Despite the great success that those existing deep learning methods already achieved, the limitations are still obvious. Particularly, these methods are imaging processing methods without considering the domain knowledge of the brain. In this paper we have designed a novel framework including the following considerations.

The first one is the individual variability of the human brain. As we know, deep learning algorithms typically include a large size of the dataset, which is the same to the brain imaging analysis. Traditionally, it is quite common to register those brains into a common template or common atlas (Tzourio-Mazoyer et al., 2002; Desikan et al., 2006; Dickie et al., 2017). In this way, all the brains can be further analyzed under the same space and with correspondence. However, it largely ignores the individual differences and all the features learned from the deep learning models are only corresponding to the common characteristics, and thus those individual differences are barely considered. The second problem is the traditional methods cannot reflect the network nature of the human brain. Traditional deep learning algorithms, such as CNN and RNN based approaches (Zou et al., 2017; Wang et al., 2019; Jiang et al., 2020), are learning the learning features through the whole brain fMRI signals (Zou et al., 2017; Wang et al., 2019) or on top of the functional/structural connectivity of whole brain voxels (ROIs) (Jiang et al., 2020). For example, the AAL atlas (Tzourio-Mazoyer et al., 2002) is frequently used to define the brain regions, and then the functional connectivity was calculated to represent the functional interactions in the brain. However, the brain is a complex dynamic network, which should be considered as the prior knowledge when designing deep learning methods to decode brain states.

To overcome the abovementioned issues and design a domain knowledge informed explainable deep learning model, in this paper, we proposed to incorporate Dense Individualized and Common Connectivity-Based Cortical Landmarks (DICCCOL) (Zhu et al., 2013) and Holistic atlases of functional networks and interactions (HAFNI) (Lv et al., 2015b; Jiang et al., 2016)-based Self-Attention Graph Pooling (DH-SAGPool) model to perform multiclassification experiments on task fMRI (t-fMRI) for characterizing and interpreting different brain states. Advantages of the proposed model is summarized as three-fold. First, DICCCOL is adopted to provide the common and consistent landmarks across the subjects and defined by group-wise consistent white-matter fiber connection patterns derived from diffusion tensor imaging (DTI) data (Zhu et al., 2013). 358 DICCCOL nodes can be successfully identified on each subject and those DICCCOL nodes retain the individual variability as well as group correspondence. Second, apart from DICCCOL system, HAFNI system is also introduced to interpret the functional brain networks (Lv et al., 2015b; Jiang et al., 2016). As we know, there has been mounting evidence that the human brain state emerges from and is realized by the interaction of multiple concurrent functional brain networks. With the dictionary learning and sparse representation (Lv et al., 2015a, 2015c) embedded in the HAFNI method, the interaction of the brain dynamics can be modeled. Combining the DICCOL and HAFNI, we consider the structural connection and functional interaction simultaneously for task classification. Thirdly, Graph Convolution Network (GCN)-based deep learning algorithm fits the nature that brain connectome is widely modeled as a large-scale complex graph. In contrast, Traditional deep learning architectures, such as CNN, only consider the influences of Euclidean based (local) neighbors, ignoring the structural connection and functional interaction of spatially remote regions (Zhang et al., 2021). In contrast to traditional CNN/RNN-based models, GCN-based models can learn features between remote brain regions, thus characterizing the fundamental non-Euclidean property of functional brain networks. GCN model has shown impressive results in modeling functional brain networks (Ktena et al., 2018; Zhang et al., 2019; Filip et al., 2020; Gadgil et al., 2020; Hu et al., 2021; Song et al., 2021; Jiang et al., 2021). However, it has not been extensively explored for brain task classification. What is more, SAGPool (Lee et al., 2019) is a self-attention graph pooling model, which is a lightweight model suitable for dealing with data with low feature dimensionality. And the model uses the top-K strategy to retain or delete graph nodes, which has some advantages in this work compared with models that also use this strategy such as DiffPool (Ying et al., 2018a), Graph U-Nets (Gao and Ji, 2019a), etc. For the DiffPool model, SAGPool has demonstrated its advantages in detail in its paper (Lee et al., 2019). For the Graph U-Nets model, its architecture is much more complex than SAGPool. In addition, the original fMRI signals of the brain have been processed by DICCCOL and HAFNI in the framework, and the features of brain states have been effectively downscaled and refined, so our work does not have a strong demand for complex model. To our best knowledge, the more the complex model we use, the more difficult we observed to investigate and explain it. Therefore, we choose the SAGPool model, which is more relevant to our experimental design, as the GCN model in our framework.

In our experiments, the human connectome project (HCP) Q1 and S1200 dataset with seven task-fMRI datasets is adopted to evaluate the effectiveness of the proposed DH-SAGPool approach. DICCCOL system is used to provide the individualized and common graph nodes across the subjects; HAFNI system helps to generate the representative features reflecting functional integration of graph nodes, and their similarities are adopted as the graph edges; SAGPool is utilized to study the input nodes and edges, and therefore generating an effective model to characterize brain states. We have reached outstanding classification accuracy when classify the task states from the HCP dataset. Additionally, importance of the nodes calculated from self-attention graph pooling layer can further help to interpret the contribution of brain nodes in the task classification.

The remaining sections are summarized as follow: in the method part, the details of generating the model of DH-SAGPool is proposed, and DICCCOL and HAFNI system are introduced to contribute to the explanation of the model; in the results part, extensive experiments are conducted to demonstrate the effectiveness of proposed DH-SAGPool. In addition, explanation of the proposed model is comprehensively discussed, and new insights are observed; in the discussion part, proposed model is explained, results are summarized, and the advantages of the proposed model are further discussed.