FLARE-AutoMSC-Val-Baseline

FLARE-AutoMSC Validation Baselines — Glioma IDH / Grade

Baseline models for MICCAI FLARE 2026 Task 2 — AutoMSC: Automated Medical Image Segmentation and Classification, on the GliomaIDHType cohort (491 multi-sequence brain MRI cases).

Each model jointly performs tumor segmentation and a case-level classification using the multi-task nnUNet framework (nnUNetCLSTrainerMTL). Both models are trained with 5-fold cross-validation and share the same images, labels, and fold splits — they differ only in the classification head. This repository contains the best-checkpoint weights (checkpoint_best.pth) for each fold plus the cross-validation evaluation report and figures.

Models

Metrics are means across the 5 CV folds (each case scored by its held-out fold). AUROC is macro for the 3-class grade model and binary for the IDH model. See each model's README.md for per-fold tables and figures.

Repository Structure

DatasetXXX_GliomaIDHType/
├── README.md                                    # per-fold results and figures (same as results.md)
├── results.md
├── figures/                                     # box plots, ROC/PRC, confusion matrix, case samples
└── nnUNetCLSTrainerMTL__nnUNetPlans__3d_fullres/
    ├── dataset.json
    ├── dataset_fingerprint.json
    ├── plans.json
    ├── eval_results/                            # per-case CSVs, fold_summary.csv, summary.json
    ├── fold_0/
    │   └── checkpoint_best.pth
    ├── ...
    └── fold_4/
        └── checkpoint_best.pth

combine_predictions.py (repo root) merges the IDH (binary) and grade (3-class) prediction CSVs into the FLARE-AutoMSC submission format (identifier, idh_wildtype, idh_mutated, grade2, grade3, grade4).

Reproduction

Cross-validation evaluation and figures are produced by fivefold_eval.py and eval-report/generate_figures.py from the AutoMSC-Baseline codebase (github.com/medfm-flare/AutoMSC-Baseline); the raw dataset provides imagesTr/, labelsTr/, splits_final.json, cls_data.csv, and cls_data_idh.csv.