Skip to content

Repository files navigation

MRI-LMICs Survey — Figure, Table & Statistical Analysis Pipeline

Analysis pipeline for: Deep Learning Super-Resolution for MRI: Technical Advances and Translational Potential for Low-Resource Settings.

The reviewer-correction pipeline regenerates the corrected analyses, promoted tables, and affected figures locally. It never publishes data or modifies GitHub.

Quick Start

# Creates the isolated environment if missing; regenerates promoted reviewer-corrected
# tables and affected figures; then runs all offline verifiers and tests.
powershell -ExecutionPolicy Bypass -File scripts/analysis/run_full_reproducibility_pipeline.ps1 -RunDate 20260817

# Supplying the private reviewer workbook additionally regenerates Fleiss'
# kappa, ordinal-weighted agreement, ICC, and the reviewer-consensus Spearman
# sensitivity analysis. The workbook is validated against the canonical title
# order and is never copied into the repository.
powershell -ExecutionPolicy Bypass -File scripts/analysis/run_full_reproducibility_pipeline.ps1 `
  -RunDate 20260909 `
  -PrivateRatingsXlsx C:\private\RECEIVED_SCORES.xlsx

Requirements

  • Python 3.11 or higher
  • A Python virtual environment (the supplied bootstrap script uses Python 3.11+)

Statistical & Geographic Equity Pipeline

The pipeline includes advanced analytics for manuscript revision:

  • Random Forest Robustness Supplement: constrained repeated held-out validation with regularized benchmarks; exploratory only.
  • Mann-Whitney U Tests: Pairwise comparison of study characteristics.
  • Fleiss' Kappa: final aggregate agreement for 48 studies scored by 11 reviewers. Individual ratings remain private.
  • Ordinal weighted agreement: supplementary multi-rater sensitivity analysis with linear and quadratic category-distance weights. Individual ratings remain private.
  • Intraclass correlation: supplementary two-way absolute-agreement ICC for single ratings and the mean of the 11 reviewers. Individual ratings remain private.
  • LMIC--TR Spearman analysis: primary study-level correlation from data/data-clean.csv and data/tr_criteria_evidence.csv, plus an optional reviewer-median sensitivity summary from the private workbook.
  • Geographic Equity: World Bank income classification mapping.

To regenerate the aggregate agreement outputs from the private workbook, run:

python scripts/analysis/statistical/run_fleiss_kappa_from_private_xlsx.py `
  --input-xlsx <private-reviewer-workbook.xlsx> `
  --output-dir tables

The command validates the 48-by-11 matrix and writes only aggregate CSV outputs; it never copies the private workbook into the repository.

To regenerate the supplementary ordinal-weighted agreement outputs from the same private workbook, run:

python scripts/analysis/statistical/run_weighted_kappa_from_private_xlsx.py `
  --input-xlsx <private-reviewer-workbook.xlsx> `
  --output-dir tables `
  --output-xlsx <private-weighted-results.xlsx>

This calculates generalized weighted Fleiss agreement for 11 reviewers and summarizes all 55 pairwise weighted Cohen kappas. The weighted results are supplementary and do not replace the prespecified standard Fleiss statistics.

To regenerate the supplementary ICC outputs from the same private workbook, run:

python scripts/analysis/statistical/run_icc_from_private_xlsx.py `
  --input-xlsx <private-reviewer-workbook.xlsx> `
  --output-dir tables `
  --output-xlsx <private-icc-results.xlsx>

The primary ICC is ICC(2,1), two-way random effects with absolute agreement; ICC(2,k) is also reported for the mean of all 11 reviewers.

To regenerate the reviewer-consensus Spearman sensitivity summary, run:

python scripts/analysis/statistical/run_lmic_tr_correlation_from_private_xlsx.py `
  --input-xlsx <private-reviewer-workbook.xlsx> `
  --output-dir tables `
  --canonical-data data/data-clean.csv

The runner requires exactly 48 papers in canonical title order and 11 complete raters for both scores. It validates LMIC scores in 1--5 and TR scores in 0--5 before writing. It then takes the per-paper median across the 11 raters and writes only tables/analysis_lmic_tr_correlation_reviewer_consensus.csv; no names, individual ratings, private path, or per-paper median is exported.

Generate Individual Outputs

# Main figures
python scripts/figures/fig1_year_distribution.py          # Figure 1: Publication Trends
python scripts/figures/fig2_architecture_distribution.py   # Figure 2: AI Architecture Landscape
python scripts/figures/fig3_lmic_relevance.py              # Figure 3: LMIC Relevance Analysis
python scripts/figures/fig4_performance_comparison.py      # Figure 4: Performance Metrics
python scripts/figures/fig5_field_strength_application.py  # Figure 5: Field Strength & Application
# Figure 6: Translational Roadmap (Manual PNG, converted to PDF by master script)

# Main tables
python scripts/tables/table1_study_characteristics.py      # Table 1: Study Characteristics
python scripts/tables/table2_ai_architectures.py           # Table 2: AI Architectures
python scripts/tables/table3_performance_metrics.py        # Table 3: Performance Metrics
python scripts/tables/table4_lmic_applicability.py         # Table 4: LMIC Applicability
python scripts/tables/table5_statistical_insights.py       # Table 5: Statistical Insights
python scripts/tables/table6_geographic_equity.py          # Table 6: Geographic Equity

Verify Installation

& .\.venv-reproducible\Scripts\python.exe -m pytest -q

The historical two-reviewer/10-study calibration is archived under provenance and is not an active result. The current aggregate agreement outputs are tables/analysis_fleiss_kappa_summary.csv and tables/analysis_fleiss_kappa_item_agreement.csv. The supplementary ordinal weighted outputs are tables/analysis_weighted_kappa_summary.csv and tables/analysis_weighted_kappa_item_agreement.csv; the supplementary ICC output is tables/analysis_icc_summary.csv. The canonical Spearman output is tables/analysis_lmic_tr_correlation.csv, and the aggregate reviewer-median sensitivity is tables/analysis_lmic_tr_correlation_reviewer_consensus.csv.

Data

Source data: data/data-clean.csv (48 primary studies; anonymized public corpus). The separate data/tr_criteria_evidence.csv file contains the final article-level evidence and binary decisions used to reproduce the TR analysis without modifying the canonical source. Reviewer identities, reviewer assignments, individual ratings, and historical calibration files remain local-only and are not part of this repository.

The public study identity contract is data/included_study_order.csv. It fixes the included corpus to canonical Paper_ID values 1–48 by title and DOI; derived tables and the private reviewer workbook are validated against this contract. Paper 24 is the Pushing the limits of low-cost ultra-low-field MRI study. IDs are never remapped by numeric shifting.

The primary LMIC--TR estimate is the canonical study-level analysis: LMIC is read from data/data-clean.csv, and TR is read from data/tr_criteria_evidence.csv. For all 48 studies, Spearman rho is 0.4059, the deterministic two-sided 10,000-permutation p-value is 0.00330, and the 10,000-bootstrap percentile 95% CI is 0.1639 to 0.6065 (seed 42; average ranks for ties). The separate reviewer-median sensitivity gives rho -0.2577, p = 0.07869, and a 95% CI of -0.5397 to 0.0397. Weighting robustness is not scorer-dependence robustness: changing TR criterion weights does not test whether the association changes when scores come from independent reviewers.

Corrected dataset refined from an initial pool of 183 papers (2020-2025).

Key Findings

Metric Value
Papers included (Primary Studies) 48
Brain MRI (dominant area) 24 (50.0%)
CNN (most common architecture) 23 (47.9%)
Low-field MRI mentioned 14 (29.2%)
High LMIC relevance (Score 4-5) 19 (39.6%)
Clinical validation reported 19 (39.6%)
Code publicly available 6 (12.5%)
Median PSNR 32.6 dB
Median SSIM 0.917
LMIC Fleiss' Kappa (11 reviewers, 48 studies) 0.505
TR Fleiss' Kappa (11 reviewers, 48 studies) 0.223

More Information

About

No description, website, or topics provided.

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages