Explainable Radiomics For Postoperative Glioma Surveillance

This repository implements a postoperative glioma surveillance pipeline on MU-Glioma-Post and compares a paper-style radiomics baseline against improved hybrid models.

Motivation

The goals were:

• reproduce the published postoperative surveillance workflow
• keep the model interpretable
• improve held-out ROC AUC with a cleaner prediction target and better features

Process

The main entrypoint is main.py, and the model-training loader lives in radiomics_pipeline/training/dataloader.py.

1. audited and indexed the longitudinal postoperative MRI dataset
2. built lesion masks from labels 1/2/3 and excluded resection-cavity-only regions
3. applied N4 bias correction and lesion-mask z-score normalization
4. extracted PyRadiomics features from T1, T1c, T2, and FLAIR
5. evaluated models with patient-held-out splits
6. narrowed the imaging backbone to T1c + FLAIR
7. moved from a progression-state framing to a forward-prediction framing
8. added curated molecular and basic clinical covariates to form a hybrid explainable-radiomics model

Comparative Results

Setting	Model	Inputs	Held-out design	ROC AUC
Christodoulou et al. (baseline paper)	LightGBM-256	Radiomics, postoperative surveillance	30 patients / 96 scans	0.80
Naive initial replication	LightGBM-64	Radiomics only, paper-style replication	30 patients / 96 scans	0.599
Paper-style hybrid attempt	LogReg-48	T1c + FLAIR radiomics + molecular/basic clinical features	30 patients / 96 scans	0.621
Forward radiomics-only	LogReg-32	T1c + FLAIR radiomics	30 patients / 84 scans	0.674
Earliest-scan hybrid screen	LogReg-48	T1c + FLAIR radiomics + molecular/basic clinical features	30 patients / 30 scans	0.873
Calibrated forward hybrid	LogReg-32	T1c + FLAIR radiomics + molecular/basic clinical features	30 patients / 84 scans	0.804

Interpretation

• Radiomics-only replication did not recover the target performance.
• Hybrid gains were strongest in the calibrated forward logistic-regression run and the earliest-scan screen, but not on the looser post_progression paper-style split.
• T1c + FLAIR was the strongest imaging backbone.
• Logistic regression was the most stable tabular model in held-out evaluation.
• The best gains came from adding age, sex, and curated molecular features to the radiomics table.

Citations

• Christodoulou RC, Vamvouras G, Pitsillos R, Solomou EE, Georgiou MF. Explainable radiomics with probability calibration for postoperative glioblastoma surveillance. European Journal of Radiology Artificial Intelligence. 2026;5:100074. doi: 10.1016/j.ejrai.2026.100074
• Mahmoud E, Gass J, Dhemesh Y, et al. MU-Glioma Post: A comprehensive dataset of automated MR multi-sequence segmentation and clinical features. Scientific Data. 2025;12:1847. doi: 10.1038/s41597-025-06011-7
• Yaseen D, Garrett F, Gass J, et al. University of Missouri Post-operative Glioma Dataset (MU-Glioma-Post) (Version 1). The Cancer Imaging Archive. 2025. doi: 10.7937/7K9K-3C83

Layout

• scripts/prep-data.sh builds the manifests and processed inputs.
• scripts/run.sh runs the calibrated forward hybrid training flow in one go.
• requirements.txt pins the Python packages used by the checked-in workflow.
• main.py is the top-level CLI entrypoint.
• radiomics_tools/metrics/ contains the reusable engineered metric helpers.
• radiomics_pipeline/training/ contains the model-side data loading helpers.
• radiomics_pipeline/workflows/ contains the Python workflow code behind the shell wrappers.
• models/calibrated/ contains the exported calibrated bundle.
• configs/ contains the PyRadiomics configuration.
• tests/ contains the metric unit tests.

This repository contains code and the retained model artifacts, not the raw TCIA data. Processed derivatives are available at Hugging Face. The recommended direction is the calibrated forward hybrid logistic-regression model: T1c + FLAIR radiomics plus curated molecular/basic clinical features.

Environment

The shell wrappers prefer ./.venv/bin/python when that virtualenv exists and otherwise fall back to python3.

Install the workflow dependencies with:

python3 -m venv .venv
./.venv/bin/python -m ensurepip --upgrade
./.venv/bin/python -m pip install -r requirements.txt

prep-data additionally requires the raw MU-Glioma-Post TCIA folder tree and the clinical workbook at the paths shown by python main.py prep-data --help.