Development of a Personalized Color Vision Correction Display Filter for Individuals with Color Vision Deficiency Using fMRI-Based Neural Responses
Decoding and characterizing color perception in color vision deficiency (CVD) using fMRI-based forward encoding models and Shared Response Model (SRM) group comparison
- Overview
- Research Questions
- Current Status
- Project Phases
- Future Directions
- Installation
- Usage
- Project Structure
- Data
- References
Can individuals with CVD distinguish colors neurally despite retinal deficits, as measured by fMRI decoding accuracy in visual cortex?
망막 결함에도 불구하고 색맹자가 신경 수준에서 색을 구별할 수 있는가? (fMRI 디코딩 정확도 측정)
- ✅ Answer: Yes! All CVD participants showed successful color decoding
- 답변: 예! 모든 색맹 참가자가 성공적인 색 디코딩을 보임
- Mean decoding accuracy: 0.592 ± 0.121 across 40 subject-ROI pairs (C010 + Procrustes pipeline)
- hV4 shows strongest selectivity: 0.613 ± 0.092 (decoding), 0.541 ± 0.283 (RDM correlation)
- RDM correlation after Procrustes: 0.381 ± 0.278 (100% positive pairs, up from 52.5% pre-alignment)
- CVD group: 0.684 ± 0.094 (numerically higher than HC 0.552 ± 0.111)
- Supporting evidence: Noise ceiling utilization 79.4% (C010 pipeline); all subjects p < 0.001 vs chance
Does CVD show inter-individual heterogeneity in neural color representations, necessitating personalized approaches?
색맹은 신경 색 표상에서 개인 간 이질성을 보이는가? 개인화된 접근이 필요한가?
- ✅ Answer: Yes, substantial individual heterogeneity confirmed via SRM group comparison
- 답변: 예, SRM 그룹 비교를 통해 상당한 개인차 확인
- SRM LOO-consistent group disparity (HC-only SRM, LOO references):
- V1: p=0.062 (g=1.16), V2: p=0.075 (g=1.04) — trending with large effects
- V3/hV4: not significant
- Individual CVD tests (Crawford & Howell 1998):
- sub-09 (protan): V1 p=0.007* — early visual cortex disruption
- sub-08 (deutan): V2 p=0.040* — mid-level visual processing impact
- sub-10 (deutan): falls within HC range — functionally normal representations
- LOSO color-dependency (CVD color-specific, HC color-agnostic):
- CVD: V2 p=0.010, V3 p=0.000, hV4 p=0.016
- HC: p=0.21–0.36 (not significant)
- CVD heterogeneity: 1.4–1.6× more dispersed than HC across all ROIs
- Key finding: Identical genotype (deuteranopia) → Opposite neural phenotypes (sub-08 vs sub-10)
- Robustness: Convergent validity with SRM-independent metrics (crossnobis r=0.486**, PCA r=0.742***)
Can three-dimensional neural profiles (magnitude, sign, structure) inform individual-specific display filter design?
3차원 신경 프로파일(크기, 부호, 구조)이 개인별 맞춤형 디스플레이 필터 설계에 활용될 수 있는가?
- 🔬 In Progress — Forward model complete, filter optimization planned
- 결과: Forward encoding model 완료, 필터 최적화 설계 중
- Forward model (Future Phase 1) validated: hV4 interpolation confirmed (perm p=0.026)
- Cool/S-axis distortion characterized as Phase 2 filter target
- LOSO zero-shot transfer validated: group prior = viable prediction engine
- Next: Stimulus-space filter optimization (Future Phase 2) using cone-shift analysis
These questions address methodological foundations for the neural-guided filter development pipeline.
Can a common color decoder be successfully applied across HC and CVD participants after alignment?
- ✅ Answer: Yes — alignment is essential; optimal pipeline is task-dependent
- LORO classification: LDA+SRM best (acc_45 = 0.793, ICC = 0.666); resolves Procrustes LDA reliability paradox
- LOCO interpolation: FE+Procrustes best (HC MAE 75.7°); ForwardEncoding is the only model with interpolation ability
- Without alignment: ALL models perform at chance (~37–39% LORO, ~90° LOCO)
- HC ≈ CVD performance (LDA: HC 0.805, CVD 0.859) → shared voxel-color mapping confirmed
- Group prior: HC-mean W improves CVD LOCO by +4–8% (leakage-free nested CV)
Can a channel-based forward encoding model predict brain responses for any hue angle in 360° circular space?
- ✅ Complete (Future Phase 1)
- Encoder: Ridge regression with GCV (ridge_gcv) — direct fit, no group prior
- Gate result: hV4 = PRIMARY GO (perm p=0.026, FE-3); V3 = CONDITIONAL (p=0.045); V1/V2 = FAIL
- Omnibus: Stouffer Z=2.87, p=0.002 — cortex-level color interpolation confirmed
- Key finding: Discrimination ≠ Interpolation — V1/V2 classify but cannot interpolate
- HC-CVD gap: Cool/S-axis (blue, purple, magenta) persistent deficit; warm gap = model-specification artifact
- LOSO Zero-Shot: hV4 ZS ≈ LORO (p=0.913) — group prior is a valid prediction engine for Phase 2
Can optimization-based filter discovery find display colors that make CVD brain responses match HC responses?
- 🎯 Planned (Future Phase 2)
- Goal: For each original color θ_orig, optimize display color θ_display via stimulus-space filter T_ψ
- Inputs from SRQ3: Target color range [180°, 315°] (cool/S-axis), per-subject optimal K, distortion direction, pairwise geometry
- Pre-validation: SRM pair-level analysis complete — sub-08 FDR 32 pairs, sub-09 FDR 7 pairs, sub-10 FDR 0 pairs (compensated)
- Participants: 10 subjects (7 HC, 3 CVD: 2 deuteranopia, 1 protanomaly)
- Paradigm: Rapid serial visual presentation (RSVP) of 8 isoluminant colors
- Runs: 6 runs per subject
- ROIs: V1, V2, V3, hV4 (defined using Wang et al. 2015 probabilistic atlas)
- Space: MNI152NLin2009cAsym, res-2
- Preprocessing: fMRIPrep v23.2.3 with MI-based coregistration
- Pipeline: C010 (2nd-level drift removal) + Procrustes alignment (validated 2026-02-09)
- Analysis methods:
- Forward encoding models (Brouwer & Heeger, 2009)
- Shared Response Model (SRM; BrainIAK) for between-subject alignment
- Crawford & Howell (1998) modified t-test for single-case inference
- Crossnobis distance (Walther et al., 2016) for SRM-independent validation
- Phase 1: C010 + Procrustes baseline decoding (validated 2026-02-09)
- Mean decoding accuracy 0.592, RDM correlation 0.381, noise ceiling utilization 79.4%
- Phase 2: SRM between-subject group comparison (HC-only SRM, LOO-consistent)
- Group: V1 p=0.062, V2 p=0.075; Individual: sub-09 V1 p=0.007*, sub-08 V2 p=0.040*
- LOSO color-dependency: CVD V2/V3/hV4 significant, HC not significant
- Phase 2b: Decoder model comparison and cross-validation (21/21 validations complete)
- Task-dependent optimality: LDA+SRM for LORO (0.793), FE+Procrustes for LOCO (75.7°)
- Procrustes/SRM alignment essential; HC ≈ CVD; group prior validated (+4–8% CVD LOCO)
- Negative results: decoder bottleneck not improvable (Result 7), sequential/MLP dead ends (Result 10)
- Robustness triangulation (A3/A4/A5):
- A3 Variance Explained: CVD VE ≥ HC (V2 g=−1.68)
- A4 Crossnobis RDM: SRM-independent convergent r=0.486**
- A5 PCA-CCA replication: convergent r=0.742***
- Future Phase 1: Forward encoding model — 360° hue interpolation validated
- hV4 = PRIMARY GO (perm p=0.026); Omnibus Z=2.87, p=0.002
- ridge_gcv encoder confirmed; smooth_tikh rejected
- HC-CVD gap: cool/S-axis persistent deficit; warm gap = model artifact
- LOSO zero-shot ≈ LORO (p=0.913) — group prior validated for Phase 2
- Track A (residual biology), Track B (CVD prediction), Track C (dimensionality) — all complete or ready
- Future Phase 2: Filter optimization — stimulus-space CVD correction via cone-shift analysis
Goal: Establish baseline decoding performance and quantify color representation quality
Pipeline (C010 + Procrustes, validated 2026-02-09):
- 1st-level GLM: FIR basis (8 delays, 0–12s)
- Voxel selection: Top 50% by FIR R²
- 2nd-level GLM: 8 HRF + 8 derivative + 12 per-run drift regressors
- Procrustes alignment: runs 1–5 aligned to run 0
- Forward encoding: 6 half-wave rectified channels, LORO cross-validation
Key Results:
| ROI | N | RDM Correlation (M ± SD) | Decoding Accuracy (M ± SD) |
|---|---|---|---|
| V1 | 10 | 0.313 ± 0.215 | 0.560 ± 0.138 |
| V2 | 10 | 0.370 ± 0.256 | 0.581 ± 0.131 |
| V3 | 10 | 0.316 ± 0.328 | 0.613 ± 0.130 |
| hV4 | 10 | 0.541 ± 0.283 | 0.613 ± 0.092 |
Documents: analysis/phase1_preprocess_decoding/README.md, analysis/METHODS_RESULTS_SUMMARY_FOR_PAPER.md
Goal: Quantify HC-CVD representational differences in SRM shared space
Method: HC-only SRM (BrainIAK) with LOO-consistent disparity analysis
- SRM trained on 7 HC subjects only; CVD projected via SVD
- LOO references: HC sub-i vs mean of other 6 HC; CVD vs same LOO references
- Three bias fixes: (1) HC-only training, (2) LOO for HC, (3) same LOO refs for CVD
- Crawford & Howell (1998) for individual CVD inference
- 10,000 permutation iterations (LOO-consistent)
Canonical script: analysis/phase2_SRM_across_between/rerun_loo_consistent.py
Key Results:
| ROI | HC LOO | CVD LOO | Separation | p (perm) | Hedges' g |
|---|---|---|---|---|---|
| V1 | 0.453 | 0.590 | 0.137 | 0.062 | 1.16 |
| V2 | 0.486 | 0.606 | 0.120 | 0.075 | 1.04 |
| V3 | 0.540 | 0.564 | 0.023 | 0.395 | 0.18 |
| hV4 | 0.700 | 0.677 | −0.023 | 0.559 | −0.14 |
Individual CVD (Crawford & Howell):
- sub-09 (protan): V1 p=0.007* — early visual cortex
- sub-08 (deutan): V2 p=0.040* — mid-level processing
- sub-10 (deutan): HC range — functionally normal
Validation (12+ tests): LOSO stability (V2 7/7), split-half (V2 both halves sig), permutation, ICC, RDM consistency, alignment comparison, crossnobis, PCA-CCA, variance explained
Documents: analysis/phase2_SRM_across_between/README.md, analysis/phase2_SRM_across_between/validation/
Goal: Validate decoder assumptions — linearity, alignment necessity, group comparability, interpolation
Methods: 6 models (LDA, Ridge, ForwardEncoding, KernelRidge, SVM, MLP) compared with LORO and LOCO CV
Key Findings:
- Task-dependent optimality: LDA+SRM for LORO classification (0.793, ICC 0.666); FE+Procrustes for LOCO interpolation (75.7°)
- Alignment is essential: Without it, ALL models perform at chance (~37–39% LORO, ~90° LOCO)
- HC ≈ CVD: Shared voxel-color mapping confirmed (justifies filter learning)
- ForwardEncoding is the only model with interpolation ability (LOCO MAE 72–83° vs 90° chance)
- LDA+SRM = optimal LORO pipeline: SRM resolves LDA fold-instability (ICC 0.013 → 0.666)
- Group prior validated: HC-mean W improves CVD LOCO by +4–8% (leakage-free nested CV)
- Cross-subject generalization: HC→CVD = HC→HC in SRM space (no group bias)
- Negative results: Decoder bottleneck not improvable (4 alt. methods all worse); sequential/MLP dead ends
Documents: analysis/phase3_decoder_comparing/model_comparison_validation/, analysis/METHODS_RESULTS_SUMMARY_FOR_PAPER.md
Our filter development follows a 2-phase approach: forward model validation (complete) → stimulus-space filter optimization (planned).
Goal: Validate continuous hue encoding — can we predict brain responses for colors not in the training set?
Method: Channel-based forward encoding Y_s = W_s @ C(θ; K) with ridge_gcv encoder, Leave-One-Color-Out (LOCO) cross-validation
Key Results:
| ROI | Optimal K | Perm p (10K) | Gate |
|---|---|---|---|
| V1 | FE-2 | 0.170 | FAIL |
| V2 | FE-3 | 0.125 | FAIL |
| V3 | FE-8 | 0.045 | CONDITIONAL |
| hV4 | FE-3 | 0.026 | PRIMARY GO |
- Omnibus: Stouffer Z=2.87, p=0.002 — cortex-level interpolation exists
- Discrimination ≠ Interpolation: V1/V2 discriminate but cannot interpolate (confirmed across all basis types including opponent)
- HC-CVD gap: Warm-color gap = model-specification artifact (>100% reduction at optimal K); Cool/S-axis gap persists at 65% → Phase 2 filter target
- LOSO Zero-Shot: hV4 ZS ≈ LORO (p=0.913) — group prior validated as prediction engine
- Track B (CVD Prediction): Per-subject K* adopted (sub-08 K*=8 → 6.4× LOCO improvement); anisotropic basis and hierarchical FE rejected
Documents: analysis/future_phase1_forward_model/README.md, analysis/future_phase1_forward_model/RESULTS.md
Goal: For each original color, find the optimal display color that makes CVD brain responses match HC responses via stimulus-space filter T_ψ
Core Innovation: Stimulus-space correction guided by cone-shift analysis (not direct voxel transformation)
Mathematical Framework:
T_ψ: θ → θ' = θ + ψ(θ)
where ψ(θ) = Σ_k [a_k sin(kθ) + b_k cos(kθ)] (Fourier parameterization)
Pre-Validation Complete: SRM pair-level distortion characterized per CVD subject
- sub-08 (deutan): 32 FDR-significant pairs, split-half r=0.73–0.84, cone-shift prediction 7/7 match
- sub-09 (protan): 7 FDR pairs (V1 magenta axis), dual dissociation with sub-08
- sub-10 (deutan, compensated): 0 FDR pairs, HC-like — cortical compensation case study
Documents: analysis/future_phase2_filter_optimization/pre_validation/notion_prevalidation.md
Original Phase Plans (Click to expand)
These were the original phase plans, now superseded by the SRM-based approach (Phase 2) and the forward model + filter pipeline above.
- Goal: Trial-aligned GPA for HC common space
- Status: Superseded by SRM (Phase 2) — archived in
analysis/archive/future_phase1_hyperalignment/
- Goal: Learn linear transformations to map CVD patterns to HC-like patterns
- Status: Superseded by SRM group comparison approach
- Goal: Learn explicit stimulus → brain mapping
- Status: Completed as Future Phase 1 (forward encoding model)
- Goal: Compute stimulus-level color corrections
- Status: Evolved into Future Phase 2 stimulus-space filter optimization
- Goal: End-to-end neural network for CVD color correction
- Status: Deferred — nonlinear models showed no benefit in current data/task
- Python 3.8+
- Conda (recommended)
- SLURM cluster (for large-scale analysis)
-
Clone repository
git clone https://github.com/yourusername/colorBlind_analysis.git cd colorBlind_analysis -
Create conda environment
conda env create -f environment.yml # Server: conda activate nilearn # Local (SRM analysis): conda activate srm
-
Verify installation
python -c "import nilearn, nibabel, sklearn; print('Success!')"
# Run forward encoding model for a single subject-ROI
conda activate nilearn
python analysis/phase1_preprocess_decoding/fir_reconstruction_BH2009_system_clean.py \
--subject 02 \
--roi V1 \
--dataset full_dataset_C010# Run canonical LOO-consistent SRM analysis
conda activate srm
mpirun -np 1 python analysis/phase2_SRM_across_between/rerun_loo_consistent.py# Run model comparison (LORO + LOCO)
python analysis/phase3_decoder_comparing/model_comparison_validation/scripts/run_comparison.pycolorBlind_analysis/
├── README.md # This file
├── CLAUDE.md # Development guide for Claude Code
│
├── analysis/ # Core analysis code
│ ├── README.md # Master analysis overview
│ ├── METHODS_RESULTS_SUMMARY_FOR_PAPER.md # Exact statistics for all phases
│ ├── filter_design_plan.md # Filter design planning
│ │
│ ├── prep_trials/ # Registration quality comparison
│ ├── roi_masks/ # ROI mask files
│ │
│ ├── phase1_preprocess_decoding/ # Phase 1: Baseline (C010 + Procrustes) ✅
│ │ ├── fir_reconstruction_BH2009_system_clean.py # Main analysis script
│ │ └── results/full_dataset_C010/ # Per-subject per-ROI results
│ │
│ ├── phase2_SRM_across_between/ # Phase 2: SRM group comparison ✅
│ │ ├── rerun_loo_consistent.py # Canonical LOO-consistent analysis
│ │ ├── validation/ # 12+ validation tests (A3/A4/A5, 1A-2D)
│ │ └── results/
│ │
│ ├── phase3_decoder_comparing/ # Phase 2b: Decoder cross-validation ✅
│ │ ├── model_comparison_validation/ # 6-model LORO + LOCO comparison
│ │ └── results/
│ │
│ ├── phase2_procrustes_cvd_hc/ # Legacy: Procrustes-based comparison
│ ├── archive/phase3_procrustes_filter/ # Legacy: Exploratory filter learning
│ │
│ ├── archive/future_phase1_hyperalignment/ # Archived: HC hyperalignment (superseded by SRM)
│ ├── future_phase1_forward_model/ # SRQ3: 360° encoder ✅ (ridge_gcv, hV4 GO)
│ ├── future_phase2_filter_optimization/ # SRQ4: Stimulus-space filter 🎯 (planned)
│ │
│ ├── comprehensive/ # Cross-phase analyses
│ ├── validation/ # Cross-pipeline validation
│ └── utils/ # Shared utilities
│
├── prediction_model_workspace/ # Legacy dev workspace (superseded)
│ ├── MASTER_PLAN.md # Original 3-phase plan (historical)
│ ├── docs/ # Early documentation + pipeline diagrams
│ └── scripts/ # Experimental scripts (migrated to analysis/)
│
├── docs/ # Documentation
│ ├── program_paper/ # Manuscript (main.tex, main_kr.tex)
│ ├── methods/ # Methodology docs
│ ├── results/ # Result summaries
│ └── technical/ # Technical reports
│
├── data/ # Local data
├── ProbAtlas_v4/ # Wang Atlas (2015)
├── papers/ # Reference papers
├── results/ # Analysis outputs (gitignored)
└── derivatives/ # Processed fMRI data (gitignored)
analysis/ — Organized by research phases (Phase 1 → Phase 2 → Phase 2b → Future Phases 1-2)
- Each phase has its own README explaining methods, results, and connections
METHODS_RESULTS_SUMMARY_FOR_PAPER.md: Authoritative source for all statistics
prediction_model_workspace/ — Legacy development workspace (superseded)
- Early experimental scripts and planning documents
- Active code has migrated to
analysis/future_phase1_forward_model/andanalysis/future_phase2_filter_optimization/
Due to participant privacy, raw fMRI data are not publicly available. Processed results and code are provided for reproducibility.
colorBlind_data/
├── sub-01/
│ ├── anat/
│ │ └── sub-01_T1w.nii.gz
│ ├── func/
│ │ ├── sub-01_task-rsvp_run-1_bold.nii.gz
│ │ ├── sub-01_task-rsvp_run-1_events.tsv
│ │ └── ... (6 runs total)
│ └── fmap/
│ └── sub-01_fieldmap.nii.gz
└── ... (sub-02 through sub-10)
- Colors: 8 isoluminant colors equally spaced in CIELAB (L*=70)
- Presentation: RSVP at 2 Hz (500ms/stimulus)
- Runs: 6 runs per subject
- Duration: ~25 min per run
See materials/ for stimulus generation code.
- Jin-il Kim — Principal investigator
-
Brouwer, G. J., & Heeger, D. J. (2009). Decoding and reconstructing color from responses in human visual cortex. Journal of Neuroscience, 29(44), 13992-14003.
- Foundation for our forward encoding model approach
-
Chen, P.-H. C., Chen, J., Yeshurun, Y., Hasson, U., Haxby, J., & Ramadge, P. J. (2015). A reduced-dimension fMRI shared response model. Advances in Neural Information Processing Systems, 28, 460-468.
- BrainIAK Shared Response Model (SRM) for between-subject alignment
-
Crawford, J. R., & Howell, D. C. (1998). Comparing an individual's test score against norms derived from small samples. The Clinical Neuropsychologist, 12(4), 482-486.
- Single-case inference for individual CVD testing
-
Walther, A., Nili, H., Ejaz, N., Alink, A., Kriegeskorte, N., & Diedrichsen, J. (2016). Reliability of dissimilarity measures for multi-voxel pattern analysis. NeuroImage, 137, 188-200.
- Cross-validated Mahalanobis distance (crossnobis) for RDM computation
-
Haxby, J. V., Connolly, A. C., & Guntupalli, J. S. (2014). Decoding neural representational spaces using multivariate pattern analysis. Annual Review of Neuroscience, 37, 435-456.
- Theoretical framework for MVPA
-
Wang, L., Mruczek, R. E., Arcaro, M. J., & Bhatt, M. (2015). Probabilistic maps of visual topography in human cortex. Cerebral Cortex, 25(10), 3911-3931.
- ROI definition atlas
- See
docs/program_paper/main.texfor complete bibliography - Key papers available in
papers/directory
Last Updated: 2026-03-16