A Multi-Level Fuzzy Explainable Prototype Network for Sex Classification Across Brain Atlases

Sex differences in resting-state functional brain organization have been widely reported, yet many deep learning models prioritize predictive performance over interpretability, limiting their ability to reveal underlying neurobiological mechanisms. We propose FXP-Net, a fuzzy explainable prototype-based neural network that combines temporal convolutional modeling with fuzzy prototype learning to enable intrinsically interpretable sex classification from resting-state fMRI. FXP-Net maps windowed ROI time series into a compact latent space in which class-specific prototypes are learned and activated through graded fuzzy membership functions, allowing each prediction to be explained via similarity to a small set of interpretable prototypes rather than opaque feature activations. To enforce interpretability, we introduce a prototype-regularized training objective that encourages intra-class clustering, inter-class separation, prototype diversity, and sparse activation. Across multiple brain atlases and repeated runs, FXP-Net achieves subject-level classification performance competitive with, and in several settings statistically superior to, a strong ConvNet baseline, with peak balanced accuracy exceeding ۹۵ %. Importantly, a multi-level interpretability analysis, combining prototype-level and decision-level attributions reveals anatomically consistent patterns across Brainnetome, Glasser, and Gordon parcellations. Quantitative overlap analysis highlights regions within the dorsolateral prefrontal cortex (DLPFC), including lateral area ۹/۱۰ and area ۴۶, as demonstrating robust cross-atlas agreement and consistently high importance. These findings are stable across runs and align with prior evidence linking DLPFC organization to sex-related differences in executive control and higher-order cognition. The Python implementation of the proposed FXP‑Net architecture is publicly available at: https://github.com/Mansooreh-Pakravan/FXPNet-Model.

کلیدواژه ها:

Fuzzy prototype learning ، Explainable deep learning ، Resting-state fMRI ، Sex differences ، Multi-atlas brain analysis