Quality-Aware Fuzzy Min-Max Neural Networks for Dynamic Brain Network Analysis

Tao Hou, Jiashuang Huang, Shu Jiang, Weiping Ding*

MOTIVATION

• Factors such as head motion, and heartbeat from the scanned volunteers impact the quality of data acquisition during the collection of raw rs-fMRI data, which leads to the generation of fuzzy information.

• The trustworthiness varies for each window due to the inconsistent data quality across different windows, leading to the uncertainty generated from the ensemble information of multiple windows.

INNOVATION

• The multi-view fuzzy min-max neural networks (MFMM) is designed under the multi-view learning framework. MFMM utilizes evidence as input patterns to generate hyperboxes based on fuzzy sets for obtaining class nodes, effectively dealing with the fuzzy information in dFCs.

• A quality-aware ensemble module is designed to deal with the uncertainty from different views with inconsistent data quality by identifying the trustworthiness of each view. D-S theory is utilized to directly model uncertainty for multiple views, enabling a more effective assessment of the quality of each view.

METHOD

• The MFMM utilizes evidence as input patterns to generate fuzzy hyperboxes, obtaining class nodes under the framework of multi-view learning. To capture multi-view features of brain networks and obtain evidence, we introduce the E2E, E2N, and N2G convolution filters.

• Dynamic quality-aware weighting is designed to evaluate the trustworthiness of each view based on uncertainty and belief. The dynamic quality-aware weighting assigns quality-aware weights to the classification results for each view. It is worth noting that the weights for each sample are different, i.e., the weights are dynamically generated.

Fig 1. Illustration of Quality-Aware Fuzzy Min-Max Neural Networks. Evidence is the input pattern of MFMM to obtain the class nodes. The quality-aware ensemble module connects evidence with Dirichlet distribution to evaluate quality-aware weighting and integrates multiple views with weighted fusion.

EXPERIMENTAL RESULTS

Fig 2. Classification results of MFMM with high (a) and low (b) quality. The size of the points (c) is associated with the average dynamic quality-aware weight of all samples, and the larger points correspond to higher quality, and vice versa.

Fig 3. The first row shows the discriminative connections, and the second row illustrates the map of the brain regions corresponding to the strongest connections.

Table 1. Performance comparison of the proposed and competing methods.