Augmenting radiological assessment of imaging evident dementias with radiomic analysis

This work is published in npj Dementia (https://doi.org/10.1038/s44400-025-00031-1).

Overview

This repository provides a comprehensive pipeline for preprocessing, feature extraction, modeling, and evaluation of neuroradiology imaging data. It integrates FreeSurfer-based volumetric analysis, white matter hyperintensity (WMH) burden estimation, and machine learning models (XGBoost) for Alzheimer’s disease (AD) and Other Imaging Evident Dementia (OIED) predictions.

Table of Contents

• Environment Setup
• Data Structure
• Preprocessing
• Model Training
• Results & Analysis
• Utilities

Environment Setup

1. Create a Conda environment:

   conda env create -f environment.yml
   # or for macOS specific
   conda env create -f environment-macos.yml

2. Activate the environment:

   conda activate neurorad-radiomics

Data Structure

• data/
  • ml_data/: preprocessed tabular datasets (ml_data_filtered.csv, ml_test_data_filtered.csv, etc.)
  • dev-model/: saved XGBoost models and tuning parameters
  • feature_config.json: definitions of volumetric, WMH, imaging, and demographic features
  • lobe_mapping.json, feature_names_map.json, etc.

Preprocessing

Scripts under preprocessing/ handle MRI data conversion and feature extraction:

• fs_scripts.py: filter and convert FreeSurfer outputs, parse volume statistics
• lst_scripts.py: run LST (Lesion Segmentation Tool) for WMH estimation
• NACC_mris.py & prep_array_job_csv_.py: generate batch arguments for array jobs on NACC datasets

Model Training

Training scripts are in train/:

• model.py: preprocessing and label generation
• wandb_sweep.py: hyperparameter sweep setup with Weights & Biases
• cv.py: cross-validation with GroupKFold
• final_train.py: train final AD, OIED models with hyperparmeters defined from sweeps

Results & Analysis

Misc. analyses and plotting scripts are in results/

Utilities

• utils/load_data.py: load and derive feature lists from JSON config
• utils/dump_dkt_atlas.py: export region names grouped by lobes