Findings-Aware Head CT Report Generation - MICCAI 2026 HEADLINE Challenge

Findings-aware retrieval and reinforcement-tuned generation for free-text reporting of non-contrast head CT.

Team IUCompPath - Sanket Kachole, Spyridon Bakas (Division of Computational Pathology, Indiana University School of Medicine). Contact: Spyridon Bakas (spbakas@iu.edu).

Challenge: https://headline26.grand-challenge.org/ · Data: https://headline.ccibonn.ai

---

Method in brief

A 3D CNN is trained on weak labels mined from the reports (haemorrhage, infarct, midline shift) and then frozen. Its 256-d bottleneck is a finding-aware embedding used to retrieve radiologist-authored reports; its spatial feature map conditions a LoRA-adapted Qwen2.5-1.5B, further tuned with GRPO against the challenge metric. The generated report is added to the retrieved pool and the most central candidate is selected by BERTScore consensus. Full description in the paper.

Official validation: a composite score of 0.2400 (retrieval-only, 500 studies), comprising a BERTScore of 0.2634, ROUGE-L of 0.2800, METEOR of 0.1697, and BLEU-4 of 0.1064.

---

1. Setup

conda create -n headline python=3.10 && conda activate headline
pip install -r requirements.txt

Training used a single NVIDIA RTX 6000 Ada (48 GB); preprocessing ran on Big Red 200. The inference container pins torch==2.4.0+cu121 for the Grand Challenge runtime (driver 535 / CUDA 12.0, T4 or A10G, 20 min per batch, no network access).

---

2. Usage paths

Path A - Inference from released weights (no training)

Download the weight bundle from Zenodo (Section 4), extract it into ./model, then build and run the container. This reproduces the submitted system.

./do_build.sh            # build the image
./do_test_run.sh         # forward pass on the sample batch in test/
./do_save.sh             # package as .tar.gz for Grand Challenge upload

inference.py reads the stacked 4D volumes from /input, runs CNN → retrieve top-80 → add one generated report → BERTScore consensus, and writes /output/radiology-text-reports.json.

Path B - Train from scratch

Run in order. Paths are argparse defaults and should be overridden per environment.

# index the corpus and make the patient-level split
python scripts/01_inventory_training_data.py
python scripts/02_build_master_index_and_split.py

# preprocess volumes + mine findings labels        (see note - HPC scripts)
#   -> headline_preprocessed/volumes/*.npz
#   -> findings label columns in outputs/reports_mining/master_index_clean.csv

# findings-aware encoder
python scripts/22_train_findings_classifier.py     # 3D CNN, 25 epochs
python scripts/23_predict_findings_all.py          # 256-d embedding per study

# retrieval + consensus                             -> Table 2
python scripts/24_findings_retrieval.py
python scripts/25_findings_tuning.py

# generation
python scripts/50_train_vlm.py                     # projector + LoRA on Qwen
python scripts/50b_train_vlm_resume.py             # continue to 5 epochs total
python scripts/60_grpo.py                          # GRPO on the challenge metric

# ensemble + package                                -> Table 3
python scripts/54_ensemble_gen_retrieval.py
python scripts/56_ensemble_pool_sweep.py           # picks TOP_K = 80
python scripts/55_build_vlm_bundle.py --epoch 100 --out model

# build the validation-phase submission CSV
python scripts/28_build_official_submission.py

Scores are computed with scripts/headline_metric.py, a local copy of the official Stage-1 metric (0.70 × BERTScore-F1 + 0.30 × mean(METEOR, ROUGE-L, BLEU-4)).

Note - two preprocessing scripts are not included. Volume windowing/ resampling and findings-label mining were run on HPC; they produce headline_preprocessed/volumes/*.npz and the label columns in master_index_clean.csv. Path A does not need them (the released bundle already contains the trained CNN and embeddings). For Path B, contact the authors for these two scripts or regenerate the inputs from the method in the paper.

---

3. Script reference

Root:

File	Purpose
inference.py	Container entry point (retrieve + generate)
do_build.sh / do_save.sh / do_test_run.sh	Build, package, test the container

scripts/ - the reproduction pipeline:

File	Purpose
01_inventory_training_data.py	Survey the training corpus
02_build_master_index_and_split.py	Master index + patient-level split
22_train_findings_classifier.py	Train the 3D findings CNN
23_predict_findings_all.py	Embed all studies with the CNN
24_findings_retrieval.py	Findings-aware cosine retrieval
25_findings_tuning.py	Tune retrieval (K, centering)
50_train_vlm.py	Train projector + LoRA on Qwen
50b_train_vlm_resume.py	Resume VLM training to 5 epochs
60_grpo.py	GRPO against the challenge metric
54_ensemble_gen_retrieval.py	Add generation to pool, consensus
56_ensemble_pool_sweep.py	Choose ensemble TOP_K
55_build_vlm_bundle.py	Package the weight bundle
28_build_official_submission.py	Build the submission CSV
headline_metric.py	Local copy of the official metric

The remaining scripts in scripts/ are the ablation baselines and recorded negative results behind Table 2 and the paper's discussion (numbered by development stage). They are kept for transparency and are not needed to reproduce the main result.

---

4. Model weights (Zenodo)

DOI: 10.5281/zenodo.22065512

Bundle (~4.6 GB): findings_cnn.pt, projector.pt, vlm/ (merged fp16 Qwen + LoRA), roberta-large/ (offline BERTScore).

The retrieval pool of report text (train_reports.json, 16,468 reports) is not distributed - redistributing it would breach the data use agreement. It must be rebuilt from a separately obtained, licensed copy of the training data with scripts/30_build_model_bundle.py.

---

Caveats

• No bone window is used, so fractures are unlikely to be detected.
• Optimisation targets text overlap, not clinical correctness; the system is not suitable for unsupervised clinical use.
• Selection is the bottleneck: 0.1935 realised against a 0.2676 oracle.

---

Citation

Kachole, S., Bakas, S. Findings-Aware Retrieval and Reinforcement-Tuned Generation for Free-Text Head CT Reporting. MICCAI 2026 HEADLINE Challenge.

Acknowledgements

Container scaffolding is derived from the organisers' starter kit, https://github.com/CCI-Bonn/headline26-starter-kit.