Decoupling Parcellation from Classification: Systematic Benchmark of Fast Brain Segmentation Methods for Alzheimer's Disease Detection

Jiadao Zou; Hongyu Guo; Wei Xi

eess.IVAug 17, 2026

Decoupling Parcellation from Classification: Systematic Benchmark of Fast Brain Segmentation Methods for Alzheimer's Disease Detection

Authors: Jiadao Zou, Hongyu Guo, Wei Xi

Organizations: Midea Group (Shanghai)

Abstract

Brain parcellation and classification are typically evaluated in isolation, yet downstream AD detection performance depends on their interaction. We decouple these components and systematically benchmark fast deep learning parcellation methods (SynthSeg+, OpenMAP-T1) against the FreeSurfer (FS-HV) clinical baseline through down- stream AD classification on OASIS-1. Our factorial design evaluates three parcellation methods, two volumetry strategies (hard vs. soft), and four classifier paradigms (clinical thresholds, supervised feedforward networks, ensemble methods, and foundation models with zero/few-shot prompting), with all results quantified using BCa Bootstrap 95% confidence intervals.

Figures & tables

Figure 1: System overview of our evaluation pipeline. The workflow comprises (i) parcellation methods (FS-HV baseline, SynthSeg+, OpenMAP-T1) producing regional volumes; (ii) two volumetry strategies (hard volumetry via argmax; soft volumetry via probability summation); and (iii) classifier paradigms: clinical threshold voting, conventional supervised learning (FNN, NCA-EVA), and foundation models in zero-shot and few-shot prompting settings. We evaluate downstream AD classification performance (F1 prioritized) on a fixed test set and compare methods relative to the FS-HV baseline.

1-Vote (Liberal: AD if $\geq$ 1 criterion met)
Config	Acc	AUC	F1 [95% BCa CI]	Prec	Rec
FS-HV	0.7407	0.9282	0.6557 [0.5672, 0.7500]	0.5000	0.9524
SS-HV	0.8148	0.8440	0.6939 [0.5600, 0.8182]	0.6071	0.8095
SS-SV	0.8148	0.8400	0.6939 [0.5600, 0.8261]	0.6071	0.8095
OMT1-SV	0.2593	0.8083	0.4118 [0.4118, 0.4118]	0.2593	1.0000
2-Vote (Moderate: AD if $\geq$ 2 criteria met)

Table 1: Clinical threshold classification across parcellations. Multi-ROI voting with k/4 criteria (k

\in

{1,2,3,4}) from Bartos et al. ( Bartos et al., 2019 ) , calibrated for FS-HV. SS-HV=SynthSeg+ hard volumetry, SS-SV=SynthSeg+ soft volumetry, FS-HV=FreeSurfer hard volumetry. F1 Bootstrap 95% CIs via stratified resampling. Note: AUC is constant within each method because threshold classifiers output fixed binary decisions without confidence scores.

FNN (Feedforward Neural Network)
Classifier	Config	Acc	AUC	F1 [95% BCa CI]	Prec	Rec
	FS-HV	0.8069	0.9079	0.8259 [0.7111, 0.9004]	0.8213	0.8310
	SS-HV	0.8305	0.9306	0.8575 [0.7452, 0.9231]	0.8525	0.8631
	SS-SV	0.8203	0.9028	0.8440 [0.7391, 0.9184]	0.8351	0.8548
	OMT1-SV	0.8203	0.9317	0.8482 [0.7400, 0.9209]	0.8332	0.8702
NCA-EVA (Ensemble with NCA Feature Selection)

Table 2: Supervised classifier performance across parcellations. FNN and NCA-EVA test results. FS-HV=FreeSurfer hard volumetry, SS-HV=SynthSeg+ hard volumetry, SS-SV=SynthSeg+ soft volumetry, OMT1-SV=OpenMAP-T1 soft volumetry. F1 Bootstrap 95% CIs via stratified resampling.

Figure 2: Few-shot prompting across parcellations. LLM F1 trajectories vs. FNN baselines (dotted). Solid lines: FS-HV, SS-HV, SS-SV, OMT1-SV. Markers: square=DS, triangle=Qw. Key patterns: (1) consistent 0→2-shot gains; (2) SynthSeg+ requires fewer examples than OpenMAP-T1; (3) DS $>$ Qw; (4) diminishing returns beyond k =6. F1 Bootstrap 95% CIs shown.

Parcellation	Model	k	Acc	AUC	F1 [95% BCa CI]
FS-HV	DeepSeek-v3	0	0.7901	0.8651	0.7734 [0.6956, 0.9302]
		2	0.8272	0.9087	0.8348 [0.6977, 0.9302]
		4	0.8148	0.9008	0.7752 [0.6667, 0.9091]
		6	0.8025	0.9000	0.7686 [0.6500, 0.8889]
		8	0.8272	0.8794	0.8017 [0.6921, 0.9134]
		10	0.7901	0.8694	0.7568 [0.6341, 0.9001]

Table 3: Few-shot prompting LLM performance with full-feature input (36 ROIs + age + sex). Acc=Accuracy, AUC=Area Under ROC Curve. F1 95% BCa CIs via stratified bootstrap.

SS-HV (SynthSeg+ Hard Volumetry)
Model	Output	Acc	AUC	F1 [95% BCa CI]	Prec	Rec
DS	Binary	0.7531	0.8333	0.7387 [0.6522, 0.9049]	0.7561	0.8333
DS	Probability	0.7901	0.8778	0.7734 [0.6921, 0.9268]	0.7763	0.8583
DS	Improvement	+4.91%	+5.34%	+4.70%	+2.67%	+3.00%
Qw	Binary	0.6543	0.4417	0.3955 [0.2051, 0.5532]	0.3581	0.4417
Qw	Probability	0.6790	0.7476	0.6452 [0.5238, 0.8182]	0.6491	0.6905

Table 4: LLM Output Format Comparison. Binary vs. probability output format across three volumetry methods. F1 95% BCa CIs computed via stratified bootstrap.

Figure 3: LLM stability across trials. Box plots show F1 distributions across 5 independent trials with different few-shot selections. Left: DS. Right: Qw. Configurations grouped by k -shot (0, 2, 6, 20) for FS-HV, SS-HV, SS-SV, OMT1-SV.

FNN (Feedforward Neural Network)
Classifier	Config	Acc	AUC	F1 [95% BCa CI]	Prec	Rec
	FS-HV	0.9012	0.9282	0.8000 [0.6667, 0.9268]	0.8421	0.7619
	SS-HV	0.8889	0.9405	0.8651 [0.7686, 0.9251]	0.8472	0.8940
	SS-SV	0.8519	0.9294	0.8071 [0.6904, 0.8857]	0.8071	0.8071
	OMT1-SV	0.8148	0.8552	0.7626 [0.6454, 0.8531]	0.7589	0.7667
NCA-EVA (Ensemble with NCA Feature Selection)

Table 5: Conventional ML Classifiers with Limited Input. Performance with 4 ROIs only (matching clinical threshold input). Uses identical FNN and NCA-EVA methods as Table 2 , with the only difference being input features (4 ROIs vs. 36 ROIs + age + sex). Full input results reported alongside Table 2 . F1 95% BCa CIs computed via stratified bootstrap.

Figure 4: Limited-input performance comparison. Few-shot LLM trajectories (square=DS, triangle=Qw) vs. clinical threshold baselines using identical 4-ROI input. Dotted lines: threshold F1. F1 Bootstrap 95% CIs shown.

FS-HV	DeepSeek-v3	0	0.6771	0.4941	0.5500 [0.4444, 0.6850]
		2	0.8357	0.8573	0.8482 [0.7503, 0.9209]
		4	0.7753	0.8249	0.8007 [0.7013, 0.8703]
		6	0.8206	0.8722	0.8426 [0.7452, 0.8976]
		8	0.7453	0.8135	0.7568 [0.6389, 0.8352]
		10	0.8357	0.8694	0.8554 [0.7583, 0.9112]
		20	0.7757	0.8271	0.8017 [0.7029, 0.8786]

Table 6: Few-shot prompting LLM performance with limited input (4 ROIs only). Acc=Accuracy, AUC=Area Under ROC Curve. F1 95% BCa CIs via stratified bootstrap.

Explore similar work

Beyond Anatomy: Explainable ASD Classification from rs-fMRI via Functional Parcellation and Graph Attention Networks
Mar 3, 2026Syeda Hareem Madani, Noureen Bibi, Adam Rafiq Jeraj +3Autism Spectrum Disorder Neuroimaging
Salim Habib University · VentureDive · National University of Computer and Emerging Sciences
Evidence Before Accuracy: A MRI-PET Fusion Network for Alzheimer Disease Classification with Causal Regional Validation
Sep 28, 2026Saeid Firouzi Daghigh, Saeed AyatAlzheimer's Disease Classification Explainable Medical Image Analysis
Department of Computer Engineering and Information Technology, Payame Noor University, Tehran, Iran
Anatomical Grounding and Leakage-Aware Multimodal Contrastive Learning for Alzheimer's Disease Classification from Structural MRI
Sep 14, 2026Paul-Gabriel Nicolae, Irina Georgiana MocanuData Leakage Alzheimer's Disease Classification
Computer Science and Engineering Department, Faculty of Automatic Control and Computers, National University of Science and Technology POLITEHNICA Bucharest, Romania

Mar 3, 2026J/K move · Enter open · S save

Beyond Anatomy: Explainable ASD Classification from rs-fMRI via Functional Parcellation and Graph Attention Networks

Syeda Hareem Madani, Noureen Bibi, Adam Rafiq Jeraj, Sumra Khan+2

Salim Habib University · VentureDive · National University of Computer and Emerging Sciences

Autism Spectrum Disorder Neuroimaging Explainable Medical Image Analysis Graph Attention Networks Functional Connectivity

Open paper AskPDF arXiv

Anatomical brain parcellations dominate rs-fMRI-based Autism Spectrum Disorder (ASD) classification, yet their rigid boundaries may fail to capture the idiosyncratic connectivity patterns that characterise ASD. We present a graph-based deep learning framework comparing anatomical (AAL, 116 ROIs) and functionally-derived (MSDL, 39 ROIs) parcellation strategies on the ABIDE I dataset. Our FSL preprocessing pipeline handles multi-site heterogeneity across 400 balanced subjects, with site-stratified 70/15/15 splits to prevent data leakage. Gaussian noise augmentation within training folds expands samples from 280 to 1,680. A three phase pipeline progresses from a baseline GCN with AAL (73.3% accuracy, AUC=0.74), to an optimised GCN with MSDL (84.0%, AUC=0.84), to a Graph Attention Network ensemble achieving 95.0% accuracy (AUC=0.98), outperforming all recent GNN-based benchmarks on ABIDE I. The 10.7-point gain from atlas substitution alone demonstrates that functional parcellation is the most impactful modelling decision. Gradient-based saliency and GNNExplainer analyses converge on the Posterior Cingulate Cortex and Precuneus as core Default Mode Network hubs, validating that model decisions reflect ASD neuropathology rather than acquisition artefacts. All code and datasets will be publicly released upon acceptance.

Abstract

Explore similar work

Beyond Anatomy: Explainable ASD Classification from rs-fMRI via Functional Parcellation and Graph Attention Networks

Evidence Before Accuracy: A MRI-PET Fusion Network for Alzheimer Disease Classification with Causal Regional Validation

Anatomical Grounding and Leakage-Aware Multimodal Contrastive Learning for Alzheimer's Disease Classification from Structural MRI