How iranchamika/paper2-segperturb-fm-radiomics works
This repository implements a reproducible lung-nodule imaging study, not a general-purpose application. Its principal workflow downloads LIDC-IDRI scans, selects and labels nodules, creates mask perturbations, extracts radiomics and foundation-model features, measures feature stability, and evaluates malignancy classification and size-stratified performance. The README documents the pipeline order and major outputs; sampled scripts confirm several extraction and analysis roles. Unsampled modules are represented by their documented responsibilities, without asserting internal calls beyond the documented workflow.
Main entry points: 05_extract_fm_embeddings.py FM embeddings · 04_extract_radiomics.py Radiomics 2D/3D · 07_downstream_classifier.py Malignancy classifier · utils.py Image and seed helpers
Components 19
Cohort preparation
- Cache BiomedCLIP
code/00_download_weights.py - Download CT scans
code/01_download_lidc.py - Select nodules
code/02_select_nodule_subset.py - Balance cohort
code/02b_balance_cohort.py - Perturb masks
code/03_generate_perturbations.py
Feature extraction
- Radiomics 2D/3D
code/04_extract_radiomics.py - FM embeddings
code/05_extract_fm_embeddings.py - Image and seed helpers
code/utils.py
Stability and evaluation
- Compute ICC
code/06_compute_icc.py - Malignancy classifier
code/07_downstream_classifier.py - Diameter subgroups
code/07b_subgroup_analysis.py - ICC results
results/icc_table.csv - Classifier metrics
results/downstream_metrics.csv - Stratum metrics
results/downstream_metrics_by_stratum.csv
Study outputs
- Study figures
code/08_make_figures.py - Manuscript
overleaf_project/main.tex
Other components
- Researcher
- TCIA LIDC-IDRI
- FM weights
Connections 30
- Researcher → Cache BiomedCLIP: caches weights
README.md - Cache BiomedCLIP → FM weights: downloads BiomedCLIP
README.md - Researcher → Download CT scans: starts pipeline
README.md - TCIA LIDC-IDRI → Download CT scans: provides scans
README.md - Download CT scans → Select nodules: supplies DICOMs
README.md - Select nodules → Balance cohort: writes cohort labels
README.md - Select nodules → Perturb masks: provides nodule masks
README.md - Perturb masks → Radiomics 2D/3D: supplies perturbed masks
README.md - Perturb masks → FM embeddings: supplies perturbed masks
README.md - FM weights → FM embeddings: loads cached weights
README.md - Image and seed helpers → Select nodules: supports image I/O
code/02_select_nodule_subset.py - Image and seed helpers → Perturb masks: supports image processing
code/03_generate_perturbations.py - Image and seed helpers → Radiomics 2D/3D: supports image processing
code/04_extract_radiomics.py - Image and seed helpers → FM embeddings: provides slice helpers
code/05_extract_fm_embeddings.py - Radiomics 2D/3D → Compute ICC: writes feature parquets
README.md - FM embeddings → Compute ICC: writes feature parquets
README.md - Compute ICC → ICC results: writes ICC table
README.md - Radiomics 2D/3D → Malignancy classifier: provides features
README.md - FM embeddings → Malignancy classifier: provides features
README.md - Balance cohort → Malignancy classifier: provides labels
README.md - ICC results → Malignancy classifier: filters features
README.md - Malignancy classifier → Classifier metrics: writes metrics
README.md - Radiomics 2D/3D → Diameter subgroups: provides features
README.md - FM embeddings → Diameter subgroups: provides features
README.md - Balance cohort → Diameter subgroups: provides labels
README.md - Diameter subgroups → Stratum metrics: writes metrics
README.md - ICC results → Study figures: supplies analysis data
README.md - Classifier metrics → Study figures: supplies analysis data
README.md - Stratum metrics → Study figures: supplies analysis data
README.md - Study figures → Manuscript: provides figures
README.md