PIME: Prototype-based Interpretable MCTS-Enhanced Brain Network Analysis for Disorder Diagnosis

Kunyu Zhang; Yanwu Yang; Jing Zhang; Xiangjie Shi; Shujian Yu

cs.LGFeb 24, 2026

PIME: Prototype-based Interpretable MCTS-Enhanced Brain Network Analysis for Disorder Diagnosis

Authors: Kunyu Zhang, Yanwu Yang, Jing Zhang, Xiangjie Shi, Shujian Yu

Organizations: Zhengzhou University, Zhengzhou, China · University Hospital Tübingen, Tübingen, Germany · China University of Geosciences, Wuhan, China · University of Science and Technology Beijing, Beijing, China ·

Abstract

Recent deep learning methods for fMRI-based diagnosis have achieved promising accuracy by modeling functional connectivity networks. However, standard approaches often struggle with noisy interactions, and conventional post-hoc attribution methods may lack reliability, potentially highlighting dataset-specific artifacts. To address these challenges, we introduce PIME, an interpretable framework that bridges intrinsic interpretability with minimal-sufficient subgraph optimization by integrating prototype-based classification and consistency training with structural perturbations during learning. This encourages a structured latent space and enables Monte Carlo Tree Search (MCTS) under a prototype-consistent objective to extract compact minimal-sufficient explanatory subgraphs post-training. Experiments on three benchmark fMRI datasets demonstrate that PIME achieves state-of-the-art performance. Furthermore, by constraining the search space via learned prototypes, PIME identifies critical brain regions that are consistent with established neuroimaging findings. Stability analysis shows 90% reproducibility and consistent explanations across atlases.

Figures & tables

Figure 1: Overview of the PIME framework. It consists of a prototype-based IB training pipeline with consistency regularization and a post-training MCTS module for explanatory subgraph extraction.

Method	ABIDE	ADNI				ADHD-200
		3-class	NC vs AD	MCI vs AD	NC vs MCI
GCN [ 1 ]	63.48 $\pm$ 7.61	58.27 $\pm$ 5.43	72.94 $\pm$ 4.72	71.63 $\pm$ 4.86	66.13 $\pm$ 5.21	63.72 $\pm$ 1.84
GAT [ 2 ]	67.92 $\pm$ 6.96	59.63 $\pm$ 5.18	72.19 $\pm$ 4.59	71.95 $\pm$ 4.71	66.28 $\pm$ 5.03	67.59 $\pm$ 1.52
GIN [ 3 ]	67.84 $\pm$ 6.79	60.12 $\pm$ 5.07	73.01 $\pm$ 4.51	72.68 $\pm$ 4.63	68.33 $\pm$ 4.89	67.81 $\pm$ 1.61
Graph Transformer [ 5 ]	69.31 $\pm$ 6.53	59.37 $\pm$ 5.26	72.74 $\pm$ 4.68	72.79 $\pm$ 4.75	68.17 $\pm$ 5.12	69.38 $\pm$ 1.31
BrainNetTF [ 4 ]	71.43 $\pm$ 4.67	65.18 $\pm$ 4.47	78.42 $\pm$ 3.83	77.56 $\pm$ 4.01	72.89 $\pm$ 4.23	71.28 $\pm$ 1.21

Table 1: Comparison with baselines on fMRI-based diagnosis. All values represent classification accuracy (%). All shows mean

\pm

std over leave-one-site-out or 5-fold CV. ADNI and ADHD-200 use 5-fold CV. ADNI tasks: 3-class (NC/MCI/AD classification), NC vs AD, MCI vs AD, and NC vs MCI. Best in bold, second-best underlined.

Dataset	w/o IB	w/o Proto	w/o Cons	w/o Sparse	w/o Div	Baseline	Full
ABIDE	70.10 $\pm$ 5.28	61.72 $\pm$ 6.43	67.54 $\pm$ 5.67	68.86 $\pm$ 5.41	70.57 $\pm$ 5.19	63.98 $\pm$ 6.21	72.47 $\pm$ 5.12
ADNI (3-class)	63.18 $\pm$ 4.52	57.39 $\pm$ 5.67	61.52 $\pm$ 4.89	62.87 $\pm$ 4.63	64.05 $\pm$ 4.47	59.21 $\pm$ 5.41	66.73 $\pm$ 4.38
ADHD-200	69.83 $\pm$ 1.26	61.54 $\pm$ 1.89	67.28 $\pm$ 1.53	68.79 $\pm$ 1.37	70.42 $\pm$ 1.21	64.37 $\pm$ 1.76	72.16 $\pm$ 1.14

Table 2: Ablation study on key components. All values represent classification accuracy (%). Each variant removes one component while keeping others.

Figure 2: Top 10 MCTS-identified critical brain regions for each diagnostic task. (a) ADNI AD vs NC. (b) ADNI MCI vs NC. (c) ADNI AD vs MCI.

Prototype	Rank 1	Rank 2	Rank 3	Rank 4	Rank 5
P1 (NC)	DMN: Precuneus L	DMN: Precuneus R	Visual: Calcarine R	Visual: Lingual L	DMN: Cuneus L
P2 (NC)	Motor: SMA R	Motor: Precentral L	Motor: Postcentral R	Auditory: Heschl L	Control: Parietal Inf
P3 (NC)	Visual: Occipital Mid	Visual: Fusiform L	DMN: PCC	Control: Frontal Mid	Visual: Calcarine L
P8 (MCI)	Limbic: ParaHippo R	Limbic: Hippo L	Limbic: ParaHippo L	Visual: Fusiform R	Temporal: Pole Mid
P9 (MCI)	SN: Insula L	SN: Insula R	SN: ACC L	SN: ACC R	Control: Frontal Inf
P10 (MCI)	SN: ACC R	Limbic: Amygdala L	SN: Frontal Med Orb	Subcortical: Thalamus	SN: Insula L

Table 3: Top-5 activated regions per prototype (ADNI). Representative prototypes: P1-3 (NC), P8-10 (MCI).

Table 4: Stability analysis.

Brain Network Analysis

Brain network analysis models structural and functional connectivity to explain brain organization, dynamics, cognition, and disease.

See all 65 papers

RegionFM: Interpretable Region-Based Brain MRI Classification Using Foundation Model Embeddings
Jul 15, 2026Wei Zhang, Ming TangMedical Imaging Foundation Models Medical Image Classification
L3S Research Center, Hannover, Germany
BrainRiem: Riemannian Prototype Learning for Source-Free Cross-Site Brain Network Diagnosis
Jun 28, 2026Kunyu Zhang, Tianxiang XuSource-Free Domain Adaptation Functional Connectivity
Zhengzhou University, Zhengzhou 450000, China · Peking University, Beijing 102600, China
Learning the Hierarchical Organization in Brain Network for Brain Disorder Diagnosis
Mar 10, 2026Jingfeng Tang, Peng Cao, Guangqi Wen +3Brain Disorder Classification Brain Network Analysis
Computer Science and Engineering, Northeastern University, Shenyang, China · School of Computer Science and Artificial Intelligence, Shandong Normal University, Jinan, China · AiShiWeiLai AI Research, Beijing, China +1

Jul 15, 2026J/K move · Enter open · S save

RegionFM: Interpretable Region-Based Brain MRI Classification Using Foundation Model Embeddings

Wei Zhang, Ming Tang

L3S Research Center, Hannover, Germany

Medical Imaging Foundation Models Medical Image Classification Explainable Medical Image Analysis Brain MRI

Open paper AskPDF arXiv

Foundation models provide powerful representations for brain MRI analysis, but their predictions remain difficult to interpret in anatomically meaningful terms. Clinical assessment of brain MRI is commonly organized around anatomically defined structures and regional abnormalities, whereas conventional explanation methods typically produce voxel- or patch-level importance maps that do not explicitly quantify the contributions of individual brain regions. To address this mismatch, we propose RegionFM, an interpretable framework that integrates anatomical segmentation with brain MRI foundation-model embeddings. RegionFM first divides each MRI scan into anatomical regions and constructs a separate MRI volume for each region. A frozen foundation model then encodes each region into an embedding, and a region-additive logistic model combines these embeddings such that every anatomical region contributes an explicit scalar term to the final prediction. This formulation supports both subject-level and cohort-level analyses of regional contributions. We evaluate RegionFM on cognitive-impairment classification using embeddings from multiple pretrained brain MRI foundation models. The results show that RegionFM maintains performance comparable to less interpretable fine-tuning approaches while providing anatomically grounded explanations. Randomized embedding ablations yield near-chance performance, indicating that the predictions rely on meaningful structure captured by the foundation-model embeddings rather than simple feature statistics. Overall, RegionFM better aligns model explanations with anatomy-based clinical reasoning while maintaining competitive predictive performance.

Abstract

Brain Network Analysis

Explore similar work

RegionFM: Interpretable Region-Based Brain MRI Classification Using Foundation Model Embeddings

BrainRiem: Riemannian Prototype Learning for Source-Free Cross-Site Brain Network Diagnosis

Learning the Hierarchical Organization in Brain Network for Brain Disorder Diagnosis