Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging
Paper • 2510.20639 • Published
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('validation'): ('csv', {}), NamedSplit('test'): ('json', {})}
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
🩺 BTB3D: Automated Radiology Report Generation on CT-RATE (500 Series Evaluation)
Official benchmark evaluation and generated report dataset using BTB3D (Better Tokens for Better 3D, NeurIPS 2025) on 500 randomly sampled 3D Chest CT series from the CT-RATE validation set.
- Model Architecture: 3D MAGViT-2 Tokenizer (16x16x8, 3rd stage) + LLaVA-LLaMA-3.1-8B-Instruct (LoRA Checkpoint-38000).
- Reference Dataset: ibrahimhamamci/CT-RATE
- Model Checkpoints: Forithmuscom/BTB3D
- Hardware: 3x NVIDIA A100-SXM4-40GB GPUs in parallel (~68 minutes total runtime).
📊 Benchmark NLP Evaluation Metrics (500 Series)
Evaluated against ground-truth human radiology reports (Findings_EN + Impressions_EN):
| Evaluation Metric | BTB3D (500 Scans) | CT2Rep Baseline |
|---|---|---|
| BLEU-1 | 33.18% (0.3318) |
~18.4% |
| BLEU-2 | 22.26% (0.2226) |
~11.2% |
| BLEU-3 | 15.64% (0.1564) |
~7.1% |
| BLEU-4 | 11.40% (0.1140) |
~4.6% |
| ROUGE-1 F1 | 46.65% (0.4665) |
~31.2% |
| ROUGE-1 Precision | 51.86% (0.5186) |
~34.5% |
| ROUGE-1 Recall | 47.32% (0.4732) |
~29.0% |
| ROUGE-2 F1 | 22.20% (0.2220) |
~11.8% |
| ROUGE-L F1 | 28.31% (0.2831) |
~21.5% |
| METEOR | 0.2859 | ~0.192 |
🫁 Clinical Pathology Concordance & Recall (500 Series)
| Pathology / Finding | Ground Truth Positive | Generated Positive | Sensitivity (Recall) | Specificity | F1-Score |
|---|---|---|---|---|---|
| Lymphadenopathy / Lymph Nodes | 477 | 454 | 90.78% | 8.70% | 93.02% |
| Lung Nodule / Mass / Lesion | 486 | 447 | 89.09% | 0.00% | 92.82% |
| Pleural Effusion | 462 | 443 | 88.53% | 10.53% | 90.39% |
| Ground Glass / Consolidation / Pneumonia | 349 | 340 | 69.34% | 35.10% | 70.25% |
| Calcification / Atheroma | 191 | 113 | 26.70% | 79.94% | 33.55% |
| Atelectasis | 124 | 102 | 25.00% | 81.12% | 27.43% |
| Bronchiectasis | 50 | 47 | 16.00% | 91.33% | 16.49% |
📁 Repository Contents
generated_reports_500.csv: Full table of 500 cases containing:VolumeName: CT volume filename (e.g.valid_130_a_1.nii.gz).Generated_Report: Full radiology text report generated by BTB3D.Ground_Truth_Findings: Original English radiologist findings.Ground_Truth_Impressions: Original English radiologist impression.Processing_Time_Sec: Inference and tokenization time in seconds.
evaluation_metrics.json: Exact quantitative scores (BLEU, ROUGE, METEOR).pathology_concordance_500.csv: Statistical analysis across 9 key thoracic pathologies.sampled_500_validation_reports.csv: Sampling manifest with voxel spacing, slice count, and metadata.run_btb3d_500_pipeline.py: Production-grade multi-GPU pipeline code.evaluate_reports.py: Metric evaluation script.
🚀 Citation
If you use these outputs or the BTB3D framework, please cite:
@article{hamamci2025btb3d,
title={Better Tokens for Better 3D: Advancing Vision--Language Modeling in 3D Medical Imaging},
author={Hamamci, Ibrahim Ethem and Er, Sezgin and Shit, Suprosanna and Reynaud, Hadrien and Yang, Dong and Guo, Pengfei and Edgar, Marc and Xu, Daguang and Kainz, Bernhard and Menze, Bjoern},
journal={arXiv preprint arXiv:2510.20639},
year={2025}
}
- Downloads last month
- 46