ABFR-KAN: Kolmogorov-Arnold Networks for Functional Brain Analysis

TL;DR

This work tackles atlas-based parcellation biases in resting-state fMRI FC analysis for brain disorder diagnosis by introducing ABFR-KAN, a transformer-based classifier that deploys Kolmogorov-Arnold Networks (KANs) within a ViT-style architecture. The method uses randomized, anatomy-aware anchor patches and iterative multi-scale patch sampling to generate robust, subject-specific FC representations, reducing structural bias and improving anatomical conformity. Extensive experiments on ABIDE I demonstrate that ABFR-KAN consistently surpasses state-of-the-art baselines in ASD classification, with strong single-site and cross-site generalization, and ablations confirm the necessity of combining random anchors, multi-scale sampling, and KAN-based modeling. The study also analyzes complexity and offers guidance on selecting KAN variants, highlighting practical trade-offs between accuracy and computational efficiency. Overall, ABFR-KAN advances reliable, atlas-free FC analysis and yields clinically meaningful improvements for ASD diagnosis.

Abstract

Functional connectivity (FC) analysis, a valuable tool for computer-aided brain disorder diagnosis, traditionally relies on atlas-based parcellation. However, issues relating to selection bias and a lack of regard for subject specificity can arise as a result of such parcellations. Addressing this, we propose ABFR-KAN, a transformer-based classification network that incorporates novel advanced brain function representation components with the power of Kolmogorov-Arnold Networks (KANs) to mitigate structural bias, improve anatomical conformity, and enhance the reliability of FC estimation. Extensive experiments on the ABIDE I dataset, including cross-site evaluation and ablation studies across varying model backbones and KAN configurations, demonstrate that ABFR-KAN consistently outperforms state-of-the-art baselines for autism spectrum distorder (ASD) classification. Our code is available at https://github.com/tbwa233/ABFR-KAN.

Figures (10)

• Figure 1: The general workflow of our proposed ABFR-KAN framework. In the anchor selection stage, a fixed set of reference patches is randomly sampled from grey matter (GM) regions to serve as anatomically unbiased functional anchors. In the patch sampling stage, multiple sets of brain patches are iteratively sampled at different spatial scales from each subject, and their FC to the anchors is computed to form robust, position-aware functional representations. These representations are then embedded with spatial information and processed by the classification network to produce subject-level predictions.
• Figure 2: (a) The GM mask (highlighted in orange and overlayed over the ch2bet template produced by holmes1998enhancement) from which our anchor patches are selected. (b) The baseline grid-based anchor selection process. (c) Our randomized anchor selection process reduces structural bias and enhances individual specificity.
• Figure 3: Visualization of anchor patch centers overlayed on the GM-mask (represented as a 3D point cloud of voxels). Anchors selected with a grid-based strategy (left) exhibit a rigid lattice distribution that extends into non-GM regions and fails to conform to cortical geometry. In contrast, randomly selected anchors (right) are distributed more adaptively within the GM, better following the anatomical shape of the brain.
• Figure 4: Histogram of distances from anchor patch centers to the nearest GM boundary voxel. Our proposed anchor selection method (red bars) produces anchor patches that are significantly closer to the GM surface than anchors selected based on a set grid (blue bars, the baseline). This indicates that randomly selected anchor patches reduce spatial bias and achieve stronger conformity to anatomy compared to the baseline.
• Figure 5: (a) The random patch sampling process. Observe how the size of the patches is consistent. (b) The iterative patch sampling process, where each subject is processed three times as a form of data augmentation, with patch sizes varying from 8$\times$8, 12$\times$12, and 16$\times$16.