PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jiadao Zou, Hongyu Guo, Wei Xi

arXiv:2608.16039v1eess.IVcs.AIcs.CV

TL;DR

Downstream Alzheimer’s disease detection depends on how parcellation and classification interact, but this interaction has not been systematically evaluated. The study benchmarks fast parcellations, volumetry strategies, and classifier paradigms against FreeSurfer on OASIS-1, finding that fast methods achieve comparable F1-scores with substantial speedups when paired with adaptive classifiers.

  • Problem

    The downstream impact of parcellation choice and hard versus soft volumetry on Alzheimer’s disease classification has not been systematically evaluated.

  • Method

    A factorial benchmark compares three parcellations, two volumetry strategies, and four classifier types on OASIS-1 using BCa Bootstrap 95% confidence intervals.

  • Results

    F1-scores of 0.84–0.87 for fast parcellations were comparable to FreeSurfer’s 0.72–0.87 while achieving 45–150× speedups with adaptive classifiers.

  • Takeaways & Limitations

    Fast parcellations paired with supervised classifiers offer high-performance, efficient alternatives, whereas calibrated clinical thresholds remain specific to FreeSurfer workflows.

  • Takeaways & Limitations

    Generalization is limited by OASIS-1’s homogeneous single-scanner, single-protocol data and the study’s single 60/20/20 split.

Abstract

from arXiv · show

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.

1 Introduction

This study evaluates brain parcellation through downstream Alzheimer’s disease classification rather than in isolation. It systematically compares fast parcellations, volumetry strategies, classifier paradigms, and input features against FS-HV using bootstrap uncertainty quantification.

  • Regional MRI biomarkers from brain parcellation support Alzheimer’s disease diagnosis and staging by quantifying atrophy in structures including the hippocampus, entorhinal cortex, and ventricles.
  • SynthSeg+ and OpenMAP-T1 achieve F1-scores of 0.84–0.87, compared with FS-HV’s 0.72–0.87, while processing scans 45–150× faster.SynthSeg+ requires 12 sec/scan, OpenMAP-T1 40 sec/scan, and FS-HV 30 min/scan.
  • For supervised classifiers, hard versus soft volumetry produces <1% F1 difference, while soft volumetry causes no degradation with clinical thresholds on FS-HV.
  • Few-shot prompting improves consistently from 0 to 2 shots; SynthSeg+ peaks at k=2 with F1=0.8684, whereas OpenMAP-T1 peaks at k=6 with F1=0.8714.SynthSeg+ therefore reaches its peak using 3× fewer examples than OpenMAP-T1.

2 Related Work

Prior work spans classical and fast deep-learning brain parcellation, hard versus soft volumetry, conventional volumetric-feature classifiers, and emerging foundation models for medical data. These approaches differ in computational cost, uncertainty handling, supervision requirements, and sensitivity to prompts or in-context examples.

  • Brain Parcellation Methods: Classical methods such as FreeSurfer achieve accurate parcellation but can require hours per scan and remain sensitive to protocol variation.Fast methods such as QuickNAT and FastSurfer reduce runtime but typically require curated annotated datasets and careful domain adaptation.
  • Volumetry from Probabilistic Segmentations: Hard volumetry sums integer voxel counts after argmax assignment, whereas soft volumetry weights voxel contributions by regional probabilities to preserve uncertainty information.Soft volumetry is theoretically appealing because it accounts for partial-volume effects and boundary uncertainty.
  • AD Classification from Volumetric Features: AD classification commonly uses clinical thresholds, machine-learning classifiers, logistic regression, gradient boosting, or shallow feedforward networks trained on regional volumes.Regional volumes are often augmented with TIV-normalized measures and demographic features.
  • Foundation Models for Medical Data: Foundation-model performance on structured medical data is mixed relative to specialized supervised models and depends strongly on prompt design, output format, and few-shot examples.Examples include GPT-4, Med-PaLM, Qwen, DeepSeek, and TabLLM.

3 Methods

The study uses 405 OASIS-1 subjects with publicly available FreeSurfer reference parcellations to compare clinical and rapid deep-learning segmentation through alternative volumetry and classification strategies. Downstream F1-scores are reported with BCa Bootstrap 95% confidence intervals, acknowledging uncertainty from the modest test set.

  • Dataset: The reproducible cohort comprised 405 OASIS-1 subjects, including 93 with mild-to-moderate AD and 312 cognitively normal controls.Subjects were drawn from 416 cross-sectional T1-weighted MRI scans aged 18–96 years.
  • Parcellation methods: Parcellation compared FS-HV, SynthSeg+, and OpenMAP-T1, with OpenMAP-T1 covering 280 regions and running in 40 seconds, 45× faster than FS-HV.FS-HV required approximately 30 minutes per scan, motivating faster deep-learning alternatives.
  • Volumetry: Volumetry evaluated hard argmax assignments versus soft probability-weighted assignments from probabilistic segmentations.Hard volumes are integer voxel counts, whereas soft volumes are real-valued and account for uncertainty and partial-volume effects.
  • Classification methods: Classification used a multi-ROI threshold baseline, an FNN, the NCA-EVA ensemble, and zero- or few-shot foundation-model prompting.Features included 36 bilateral anatomical regions, with age and sex additionally entering the FNN.
  • Uncertainty quantification: BCa Bootstrap 95% confidence intervals quantified every primary F1-score, with 2,000 bootstrap iterations per configuration and test-set uncertainty explicitly considered.The test set contained 81 subjects: 21 AD and 60 cognitively normal controls; CI widths ranged from 0.10 to 0.49.

4 Results

Results compare three parcellation methods, two volumetry strategies, and four classifier types using BCa Bootstrap 95% confidence intervals, with F1-score prioritized under class imbalance. Given the test sample size of n=81, interpretation emphasizes effect sizes and consistent directional patterns rather than isolated peaks.

  • Evaluation design: F1-score is prioritized for all comparisons because class imbalance can over-reward majority-class predictions, while accuracy and AUC provide secondary context.The 6:2:2 split yields train 243, validation 81, and test 81 participants, including 21 AD and 60 CN in the test set.
  • Evaluation design: The factorial evaluation spans FS-HV, SynthSeg+, and OpenMAP-T1 parcellation; hard and soft volumetry; and clinical threshold, FNN, NCA-EVA, and LLM classifiers.Results are organized by research questions and reported with BCa Bootstrap 95% CIs.
  • Interpretation: With n=81, results emphasize effect sizes and consistent directional patterns rather than isolated peak values.This interpretation is stated as a reporting principle for the benchmark.

4.1 RQ1: Parcellation Method and Volumetry Strategy Comparison

Fast deep-learning parcellation methods generally match FS-HV in downstream AD classification while reducing scan time substantially. Performance depends on classifier and prompting strategy, whereas SynthSeg+ volumetry choice has little effect under clinical thresholds.

  • Clinical thresholds: At moderate clinical voting thresholds, FS-HV outperforms SynthSeg+, but SynthSeg+ matches or exceeds FS-HV at liberal voting.At 2 votes, FS-HV F1=0.8085 versus SynthSeg+ 0.59–0.61; at 1 vote, SynthSeg+ reaches 0.6939 versus FS-HV 0.6557.
  • Clinical thresholds: SynthSeg+ hard and soft volumetry produce nearly identical F1-scores and confidence-interval widths across clinical voting schemes.The average confidence-interval width is 0.36 for both strategies.
  • Supervised classifiers: FNN F1-scores span 0.83–0.86 across parcellations, with overlapping confidence intervals and numerically comparable or higher SynthSeg+ and OpenMAP-T1 results than FS-HV.FNN scores were SS-HV 0.8575, SS-SV 0.8440, OMT1-SV 0.8482, and FS-HV 0.8259.
  • Supervised classifiers: 45–150× faster inference accompanies competitive supervised performance for SynthSeg+ and OpenMAP-T1 versus FS-HV.The reported scan times are 12 sec for SynthSeg+, 40 sec for OpenMAP-T1, and 30 min for FS-HV.
  • Few-shot prompting: Few-shot prompting improves consistently from 0 to 2 shots, with SynthSeg+ typically reaching competitive performance at k=2–4 while OpenMAP-T1 generally requires k=6–10.Performance gains diminish beyond k=6, and larger DS models outperform Qw by an average F1 advantage of +6.8%.

4.2 RQ2: Classifier Paradigm Comparison

Classifier paradigm choice strongly affects Alzheimer’s detection: supervised methods generalize better across parcellations, while LLM performance depends on output format, model size, and prompt-example selection. Multi-trial evaluation supports robust uncertainty quantification and deployment feasibility.

  • Four paradigms were compared: clinical thresholds, FNN, NCA-EVA, and few-shot prompting LLMs.
  • F1=0.66–0.81 for FS-HV-calibrated clinical thresholds, versus F1=0.37–0.61 after transfer to SynthSeg+ volumes.Clinical thresholds require zero training cost and provide instant inference, but transferability is limited.
  • Supervised classifiers achieve F1=0.81–0.86 across parcellations, with FNN at 0.83–0.86 and NCA-EVA at 0.81–0.86.Both supervised approaches offer comparable or higher performance than clinical thresholds when appropriate training data are available.
  • +43.3% average F1 improvement resulted from probability rather than binary LLM outputs across all six configurations.The average gain was +66.5% for Qwen and +20.1% for DeepSeek; the largest single DeepSeek gain was +47.4% on SS-SV.
  • Probability outputs shifted both lower and upper BCa Bootstrap 95% confidence bounds upward in all six configurations.For Qwen on SS-SV, the lower bound rose from 0.2500 to 0.6383 and the upper bound from 0.5833 to 0.8889.
  • Average F1 standard deviation was 0.0176 for DS versus 0.0498 for Qw across five trials, indicating greater DS stability.DS maintained a standard-deviation range of 0.0052-0.0326, while Qw variance depended on k-shot selection.
  • Pooling five trials (n=405) produced approximately 50% narrower confidence intervals than single-trial estimates (n=81).F1 coefficient of variation ranged from 0.68% to 15.15%, with most configurations below 10%.

4.3 Volumetry Strategy Effects

Hard and soft volumetry generally produce comparable Alzheimer’s classification performance, with supervised and few-shot models showing less than 1% average F1 difference. Ensemble methods are more sensitive to the volumetry strategy, while clinical thresholds remain compatible with both strategies on FS-HV volumes.

  • Clinical thresholds: Clinical-reference threshold classifiers achieve comparable F1-scores with hard and soft volumetry on FS-HV volumes.Clinical thresholds were calibrated using FS-HV volumes, supporting compatibility with both strategies.
  • Clinical thresholds: SynthSeg+ hard and soft volumetry yield nearly identical F1-scores and BCa confidence interval widths, averaging 0.37.This equivalence holds across all voting schemes under clinical threshold classifiers.
  • Supervised classifiers: FNNs show nearly identical F1-scores for SS-HV (0.8575) and SS-SV (0.8440), with similar BCa Bootstrap 95% CI widths of 0.1778 and 0.1793.The results indicate equivalence between hard and soft volumetry in both performance and uncertainty.
  • Ensemble methods: NCA-EVA reaches 0.8615 on SS-HV versus FNN’s 0.8575, but falls to 0.8128 on SS-SV versus FNN’s 0.8440, a -3.7% relative difference.These modest differences suggest ensemble feature selection can interact differently with probabilistic and discrete inputs.
  • Supervised and ensemble methods: Supervised classifiers show less than 1% average F1 difference between hard and soft volumetry, whereas ensemble methods vary more with feature selection.Few-shot prompting LLMs likewise show typically less than 1% F1 differences because adaptive feature transformations can accommodate either input representation.

4.4 RQ3: Input Feature Selection Effects

With only four clinical ROIs and no demographics, supervised classifiers remain strong but parcellation-dependent, while few-shot foundation models vary more with model size, examples, and parcellation. Restricted inputs particularly impair zero-shot LLMs and OpenMAP-T1, indicating greater dependence on richer feature context.

  • Limited-input supervised classification: Using four ROIs without demographics, FNN reaches F1=0.8651 with SS-HV, while limited-input performance remains parcellation-dependent.The four ROIs are bilateral hippocampi and inferior lateral ventricles; FNN and NCA-EVA retain the Table 2 architectures and training procedures.
  • Limited-input supervised classification: FNN improves over the clinical threshold by +8.1% with SS-HV and +0.9% with SS-SV, matches FS-HV, and is -4.7% lower with OMT1-SV.The clinical threshold is F1=0.8000 [95% BCa CI: 0.6667, 0.9268]. NCA-EVA exceeds it only with OMT1-SV (+2.8%).
  • Few-shot foundation models: Few-shot DeepSeek reaches F1-scores of 0.59–0.88 with k=6-20 and exceeds the clinical threshold in several limited-input configurations, but typically trails supervised baselines.Limited-input supervised ranges are FNN: 0.76–0.87 and NCA-EVA: 0.69–0.82.
  • Zero-shot limitations: Limited-input zero-shot SynthSeg+ performance ranges from 0.47 to 0.57, far below the clinical threshold of 0.8000, whereas full-feature zero-shot reaches F1=0.74-0.76.The contrast indicates that demographics and comprehensive ROI sets support better zero-shot reasoning.
  • Parcellation information requirements: OpenMAP-T1 loses 7-44% F1 with limited input across k values, indicating that its 280-region JHU atlas requires richer ROI information despite additional examples.The result reflects a granularity-efficiency trade-off: finer parcellation creates dependencies across a broader anatomical context.
  • Few-shot foundation models: DeepSeek achieves F1=0.72-0.84 with limited input, compared with Qwen at F1=0.61-0.69, and can surpass supervised methods when examples are strategically selected.Larger foundation models can leverage minimal critical anatomical information without demographic covariates.

5 Discussion

Fast deep-learning parcellations generally match or exceed FS-HV for downstream AD classification, but performance depends on classifier compatibility, parcellation granularity, prompting strategy, and volumetry representation. Clinical thresholds remain atlas-dependent, whereas supervised and few-shot methods provide broader robustness with different sample-efficiency and stability trade-offs.

  • Classifier-Dependent Parcellation Performance: SynthSeg+ and OpenMAP-T1 achieve comparable or higher F1-scores than FS-HV in supervised and few-shot settings while providing 45-150× faster inference.SynthSeg+ requires 12 sec/scan and OpenMAP-T1 40 sec/scan, compared with 30 min/scan for FS-HV.
  • Classifier-Dependent Parcellation Performance: FS-HV clinical thresholds outperform SynthSeg+ thresholds because calibration to FS-HV volumes does not transfer across atlases.FS-HV achieves F1=0.72-0.87 across voting thresholds, versus SynthSeg+ F1=0.52-0.78; FS-HV is optimal at 2-3 votes, while SynthSeg+ is optimal at 1 vote.
  • Classifier Paradigms: Supervised classifiers generalize more consistently across parcellations than fixed thresholds, while FNN and NCA-EVA show comparable overall performance with substantial confidence-interval overlap.Performance patterns depend on the parcellation–volumetry combination, but both supervised paradigms remain competitive when training data are available.
  • Parcellation Granularity and Sample Efficiency: OpenMAP-T1’s 280 regions do not consistently outperform SynthSeg+’s 100 regions, indicating that coarser parcellations can provide sufficient resolution with greater sample efficiency.With limited ROI sets, SynthSeg+ supports effective reasoning when k≥6, whereas OpenMAP-T1 degrades across all k values.
  • Few-Shot Prompting vs. Supervised Methods: Few-shot prompting competes with supervised methods using minimal labels, with most LLM configurations improving substantially from k=0 to k=2 before plateauing after k=6.Probability outputs improve F1 by +35.0% on average, with a range of -2.5% to +78.6%; DeepSeek has average F1 standard deviation 0.0176 versus 0.0498 for Qwen.
  • Volumetry Strategy: Volumetry strategy has negligible impact for supervised classifiers but more variable effects for clinical thresholds and ensemble methods.For transferred thresholds, performance degrades on SynthSeg+ regardless of hard or soft volumetry; on SS-SV, NCA-EVA reaches 0.8128 versus FNN 0.8440, a -3.7% relative difference.

6 Conclusion

This study decouples brain parcellation from classification to benchmark fast parcellations, volumetry strategies, and classifier paradigms for downstream AD detection on OASIS-1. It finds that fast parcellations can perform competitively, but classifier compatibility, feature context, and data availability shape deployment choices.

  • Study design: The factorial benchmark compares SynthSeg+, OpenMAP-T1, and FS-HV across hard versus soft volumetry and four classifier paradigms for downstream AD classification.The classifier paradigms include clinical thresholds, supervised feedforward networks, ensemble methods, and foundation models with zero/few-shot prompting.
  • Key findings: F1-scores of 0.84-0.87 for fast parcellations are comparable to FS-HV at 0.72-0.87, with 45-150× speedups using adaptive classifiers.FS-HV-calibrated clinical thresholds do not transfer reliably to other parcellations, highlighting atlas–classifier compatibility in rule-based workflows.
  • Key findings: Coarser 100-region parcellations reach competitive performance with fewer examples than finer 280-region parcellations, with minimal F1 difference.This establishes a sample-efficiency trade-off associated with parcellation granularity.
  • Key findings: With 4 ROIs, FNN achieves F1=0.76-0.87 and NCA-EVA F1=0.69-0.82, while few-shot LLMs need richer context or more examples to close the gap.The conclusion therefore favors conventional supervised methods for minimal-feature deployments.
  • Practical guidance: For labeled data, fast parcellations with FNN F1=0.83-0.86 or NCA-EVA F1=0.81-0.86 offer high performance and efficiency; few-shot LLMs are alternatives when k=2-6 examples are available.The guidance recommends FS-HV with calibrated clinical thresholds for rule-based workflows and notes that LLM probability outputs and rationales support auditing borderline cases.
  • Contribution: The study presents the first systematic benchmark of fast parcellations, volumetry strategies, and LLM-based classifiers against FS-HV baselines with bootstrap uncertainty quantification on OASIS-1.This novelty claim is stated by the authors as applying to their knowledge.

Ethics & Reproducibility

The study uses publicly available OASIS-1 data under appropriate data use agreements, collected without new human subjects research. Code, models, protocols, and data splits are available upon request and will be publicly released upon publication.

  • All experiments use publicly available OASIS-1 data under appropriate data use agreements, with no new human subjects data collected.
  • Code, trained models, evaluation protocols, and data splits are available from the corresponding author upon request and will be publicly released upon publication.
Loading 2608.16039v1…