MRF-SEEG

About this project

Project Summary

MRF-SEEG investigates whether quantitative MRI tissue properties derived from Magnetic Resonance Fingerprinting (MRF) can serve as imaging biomarkers for epileptogenic tissue localization in patients with drug-resistant epilepsy undergoing stereo-EEG (SEEG) evaluation.

The core hypothesis is that T1 and T2 relaxation times at SEEG electrode contact locations - and the correlations between those MRF values and interictal/ictal spectral EEG activity - differ systematically between the epileptogenic zone (EZ) and non-involved zones (NIZ).

Cohort: N = 13 patients with drug-resistant focal epilepsy who underwent pre-operative MRF imaging followed by SEEG implantation at Cleveland Clinic. Active patients: ith2, ith5, ith6, ith7, ith9, ith10, ith11, ith13 Related work: MRF-CLIN manuscript (accepted, 2026)

Research Questions

1. Do MRF-derived T1/T2 values differ at EZ vs. NIZ electrode contacts?
2. Do quantitative tissue properties correlate with interictal/ictal spectral EEG power at those contacts?
3. Can multimodal features (MRF + radiomics texture + spectral band power) improve seizure zone classification beyond imaging alone?
4. Does distance from the EZ explain variance in MRF or spectral signatures?

Project Timeline & Phases

Phase 1 - Data Preparation (Aug-Oct 2024)

Key scripts: MRF-SEEG-data-prep.R, R/mrf-seeg-data-prep-unified.R, merge-MRF-CLIN-data.R

• Built the core data pipeline: co-registration of post-implant CT to pre-op MRF, electrode contact extraction using POM files (contact names + MNI coordinates)
• Merged clinical annotations (EZ labels, seizure outcome, resection data) to SEEG contact table
• Validated VEP atlas correspondence for anatomical parcellation of contacts
• Established 2 mm spherical ROI around each contact as the unit of MRF sampling (optimized from 1/2/4 mm comparison; see doc/compare_1_2_4mm.qmd)
• Initial cohort: 13 patients; usable contacts from gray matter contacts only
• Output: Master dataset data/mrf-seeg-initial-13.xlsx, data/ith-all-2024-09-19.xlsx

Phase 2 - Normalization Development (Nov 2024 - Jan 2026)

Key scripts: R/2024-11-24-generate_normalized_brains.R, R/2025-11-14.R, mrf-normalization.py Documentation: doc/2025-01-07-first-look-at-mrf-values.qmd, 2026-01-22-norm_mtg.qmd

Inter-patient variability in raw MRF T1/T2 values is substantial due to scanner drift and protocol differences across acquisition dates. Multiple normalization strategies were evaluated:

Approach	Notes	Outcome
Voxel-wise Z-score vs. atlas	Blurring at gray-white junction → spurious values	Abandoned
Published lookup tables (AES 2025)	Still boundary artifacts	Partial use
Nyúl histogram matching	Good between-patient alignment	Tested
LSQ (least-squares)	Computationally stable	Tested
Fuzzy C-means tissue segmentation	Promising but complex	Tested
Patient-wise Z-score	Simple, interpretable, robust to outliers	Current approach

Breakthrough: Patient-wise Z-score normalization (T1/T2 standardized within each patient’s own brain tissue distribution) proved most practical. Explored via Python intensity_normalization package and R.

Challenges:

• Voxel-wise atlas approaches failed at tissue boundaries due to atlas blurring
• Nyúl/LSQ require careful anchor tissue selection
• No single normalization was universally superior - documented in 2026-01-22-norm_mtg.qmd

Phase 3 - SEEG Power Spectral Density Ingestion (Feb 2026)

Key scripts: R/2026-02-17-psd-first-look.R, R/2026-02-17-psd-calculations.R, R/2026-02-19-psd_graphs.R

• Ingested per-patient SEEG PSD data stored as MATLAB .mat files in HDF5 format (rhdf5 package)
• Computed frequency band power: delta (1-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), beta (12-30 Hz), gamma (30-80 Hz), high gamma (80-200 Hz)
• Generated per-contact band power summaries and merged to the master MRF-SEEG table
• First-look visualizations showed band power differences across clinical zones

Challenge: HDF5 file structure varied slightly between patients; required patient-specific parsing logic.

Phase 4 - Correlation Analyses & Statistical Modeling (Feb-Apr 2026)

Key scripts: R/2026-02-23-psd-correlations.R, R/2026-02-26-psd-correlations.R, R/2026-03-09-logistic-regression.R, R/2026-04-08-group-level-psd.R, R/2026-04-13-group-analysis-psd.R

• Computed Pearson/Spearman correlations between T1/T2 and each spectral band across contacts within patients
• Fitted generalized linear mixed-effects models (GLMM) with EZ membership as binary outcome, MRF and/or spectral features as predictors, random intercepts per patient
• AUC summaries: results/glmm_auc_summary_2026-04-13.csv
• Logistic regression with bootstrapped 95% CIs for key predictors

Breakthrough: Certain frequency bands (particularly high gamma) showed consistent positive correlation with MRF T1/T2 deviations in the EZ across patients, supporting the multimodal signal.

Challenge: Nested data structure (contacts within patients) requires careful handling; naive models inflate significance. Mixed-effects approach chosen but requires sufficient random-effects variation across N=13 patients.

Phase 5 - Radiomics Feature Extraction (Nov 2025 - Apr 2026)

Key scripts: extract_radiomics.ipynb, merge_radiomics_to_clinical.ipynb, radiomics_feature_decide.ipynb

• Extracted first-order and texture (GLCM) radiomics features from 2 mm spherical ROIs using PyRadiomics
• Features per contact: mean intensity, entropy, GLCM contrast, correlation, energy, homogeneity
• Merged to master table; feature selection via correlation filtering and PCA
• PCA attempted on combined MRF + radiomics feature space (2026-04-13-pca-attempt.R, 2026-04-13-pca-gm-only.R)

Challenge: Radiomics features are highly correlated with mean T1/T2 - limited additive value in univariate analyses. More useful in combined multivariate embedding (UMAP).

Phase 6 - Publication Figures & Final Analyses (Apr-Jun 2026)

Key scripts: R/fig2_raincloud.R, R/fig3_fingerprint.R, R/fig4_corr_spectrum.R, R/fig5_umap.R, R/2026-06-01-distance-to-EZ.R, R/2026-06-02-distance-to-EZ.R

Current focus - generating finalized publication figures and distance-to-EZ analyses:

Figure	Description	Status
Fig 1	T1/T2 distributions across clinical zones	In development
Fig 2	Raincloud plots of spectral band power by zone	Generated (R/fig2_final.png)
Fig 3	Patient fingerprint matrix (T1/T2 vs. band power correlations)	Generated (R/fig3_final.png)
Fig 4	Correlation spectrum across frequency bands	In development
Fig 5	UMAP embedding: MRF + radiomics + spectral features	Generated (R/fig5_umap_v3.R)

Distance-to-EZ analysis (most recent, Jun 2026): Testing whether MRF and spectral signatures vary as a function of spatial distance from the EZ centroid - may serve as continuous validation of the spatial specificity of the signal.

Data Architecture

Input Data

• Clinical table: data/ith-all-2025-01-22.xlsx - demographics, EZ labels, surgical outcomes
• Per-patient MRF volumes: mrf_T1__@_#fsWarped.nii.gz, mrf_T2__@_#fsWarped.nii.gz (FreeSurfer-warped to MNI)
• Electrode POM files: Contact names and MNI coordinates per patient
• SEEG PSD: .mat (HDF5) files per patient
• VEP atlas: aparc+aseg.vep.nii.gz + lookup table for anatomical parcellation

Processed / Intermediate

• data/mrf_seeg_merged_data_2025-11-14.csv - master merged dataset (~3.8 MB)
• data/mrf_psd_long_ith{N}_DATE.Rds - per-patient long-format PSD-merged data
• results/radiomics_features/ith{N}_radiomics_features.csv - per-patient PyRadiomics output

Key Output Files

• results/glmm_auc_summary_2026-04-13.csv
• results/glmm_coef_table_2026-04-13.csv
• results/t1_normalization_stats.csv, results/t2_normalization_stats.csv
• results/summary_statistics.csv
• results/table1_demographics.docx

Technical Challenges & Solutions

1. MRF Normalization

Problem: Raw MRF T1/T2 values vary substantially between patients due to scanner/protocol drift, making direct cross-patient comparisons unreliable. Attempted solutions: Voxel-wise atlas Z-score, Nyúl histogram matching, LSQ, Fuzzy C-means. Current solution: Patient-wise Z-score normalization. Each contact’s T1/T2 is expressed as standard deviations from that patient’s own tissue distribution. Residual concern: Does not account for true between-patient biological variation; treats all between-patient variance as noise.

2. Electrode Localization

Problem: Post-implant CT to pre-op MRI co-registration is complicated by brain shift post-implantation and susceptibility artifacts from the electrodes. Solution: Manual QC of each patient’s contact localization; contacts with uncertain placement excluded.

3. ROI Definition

Problem: What spatial extent should be used to sample MRF at each contact? Solution: 2 mm sphere (established via doc/compare_1_2_4mm.qmd); 1 mm too noisy, 4 mm over-samples.

4. Statistical Model Structure

Problem: Contacts are nested within patients; patients have different numbers of contacts and EZ proportions. Solution: GLMM with random intercepts per patient. N=13 limits random-effects estimation precision.

5. SEEG PSD Data Format

Problem: MATLAB HDF5 .mat files have patient-specific internal structure. Solution: Custom per-patient parsing logic in rhdf5; harmonized into long-format tidy data.

Key Findings (Preliminary)

• MRF T1/T2 values (patient-normalized) tend to be elevated at EZ contacts compared to NIZ, consistent with microstructural tissue abnormality in epileptogenic regions
• High-gamma band power shows positive correlation with T1/T2 deviation at the contact level
• Patient “fingerprint” patterns (Fig 3) are heterogeneous - some patients show strong MRF-spectral coupling, others do not - suggesting patient-level moderators (lesional vs. non-lesional, focal vs. multifocal EZ)
• UMAP of combined MRF + radiomics + spectral features shows partial separation of EZ from NIZ contacts
• Distance-to-EZ analysis (ongoing) may provide a continuous validation metric

Open Questions & Next Steps

• Finalize Fig 1 (T1/T2 zone distributions)
• Complete Fig 4 (correlation spectrum)
• Complete distance-to-EZ analysis and determine if it adds to figures/manuscript
• Finalize mixed-effects model table for manuscript
• Draft manuscript (Methods → Results → Discussion)
• Subgroup analysis: lesional vs. MRI-negative patients

Repository Structure

MRF-SEEG/
├── R/                    # 85+ analysis scripts (date-prefixed)
├── data/                 # Input and processed datasets
│   └── raw/              # Original source files
├── doc/                  # Meeting notes, exploratory QMDs
│   └── AES-2025/         # AES 2025 abstract materials
├── results/              # Model outputs, summary CSVs
│   └── radiomics_features/
├── figures/              # Publication figures and per-patient plots
├── *.ipynb               # Python: radiomics extraction, feature selection
├── mrf-normalization.py  # Python: normalization comparison tool
├── project-overview.qmd  # This file
└── 2026-01-22-norm_mtg.qmd  # Normalization methods documentation

Dependencies

R packages: tidyverse, patchwork, ggrepel, lme4, lmerTest, pROC, RNifti, neurobase, rhdf5, caret, umap, ggdist (raincloud), oro.nifti

Python packages: pyradiomics, intensity_normalization, nibabel, numpy, pandas, h5py, matplotlib

Last updated: r Sys.Date()

Concept notes and other ideas

Reading notes

Meeting notes

Weeks in which I worked on this project