EOLP/VAUBAN

Overview

Note: This model is an artificial intelligence tool capable of automatically detecting lung nodules in 3D computed tomography (CT) scans. As with all medical AI, this model remains an assistive tool: it does not replace the expertise of a radiologist and requires clinical validation for diagnostic use.

EOLP/VAUBAN is a 3D object detection model exported in ONNX format. It was trained using the state-of-the-art nnDetection framework on the public LIDC-IDRI database. The architecture is a Retina U-Net optimized for 3D medical object detection.

Through a 5-fold cross-validation scheme, the model has demonstrated state-of-the-art performance to the best of our knowledge, for instance when benchmarked against the established LUNA16 challenge. Additionally, the model was tested on a modified dataset featuring a truncated lung field of view to estimate its behavior during non-whole-lung acquisitions. This truncation showed no significant impact on the model's detection performance (results not shown).

Yet, deeper validation on independent and proprietary datasets corresponding to your particular needs is of your responsability and heavily encouraged before prod.

Pre-compiled TensorRT Cache Delivery
Along with the ONNX model weights (retinaunet_vauban.onnx), this repository provides an optional TensorRT execution cache (trt_engine_cache_fp16.zip). This cache eliminates the compilation/parsing time (which can take several minutes) during the first inference by loading a graph already optimized and compiled for the target GPU.

Use the _shaped version for TensorRT execution to avoid graph splitting and ensure maximum performance.

⚠️ Strict Compatibility Requirements: The TensorRT engine contained in this archive was compiled ahead-of-time. It will nearly only work if the target host environment matches the following configuration:

• GPU Hardware: sm86 (NVIDIA Ampere Architecture). Compatible with: RTX 3070, RTX 3080, RTX 3090, RTX A10G, RTX A40.
• Precision: FP16
• ONNX Runtime: Version 1.19.2
• TensorRT: Version 10.3.0
• CUDA: Version 12.2 (or a compatible 12.x series used by TRT 10.3 libraries)
• OS: Ubuntu 22.04 LTS x86_64

What if my system is not compatible?
If your container crashes (e.g., throwing a TensorRT engine failed to load error), or if you are using a different GPU architecture (e.g., Tesla T4, Ada Lovelace RTX 4x00): Do not use, or simply delete, the extracted files from this zip archive. Run the ONNX inference using the TensorRT backend as usual. ONNX Runtime will detect the missing or mismatched cache and will automatically recompile/re-parse a new, valid engine tailored to your machine. This takes a few minutes on the very first inference, but all subsequent ones will be accelerated.

Intended Use

• Primary Use: 3D detection and localization of pulmonary nodules in CT scans.
• Output: Bounding boxes and confidence scores for lung nodules.
• Integration: Assistive bounding box generation for radiological review, intended to be integrated into ONNX (sliding-window) inference pipelines.
• Out-of-Scope:
  • Modalities other than CT (e.g., x-ray, MRI, ultrasound).
  • Autonomous diagnosis or malignancy classification without clinical review by physicians.
  • Usage on pediatric populations without specific calibration or disclaimer.

Training Details

• Dataset: The Lung Image Database Consortium (LIDC-IDRI). Incorporates 1,018 diagnostic and lung cancer thoracic CT scans originating from 1,010 patients, with marked-up annotated lesions.
• Architecture: Retina U-Net (3D).
• Framework: Trained end-to-end using the autoML framework nnDetection. Default parameters from the nnDetection pipeline were strictly preserved to ensure optimal and self-configuring hyperparameter selection and reproducibility.
• Validation Scheme: 5-Fold cross-validation (only one fold deployed).

Fig. 1 - FROC analysis of the model (5-Fold cross-validation)

Evaluation & Performance

To the best of our knowledge, this model achieves state-of-the-art results, for instance when benchmarked against the LUNA16 challenge dataset criteria. Keep in mind that performance of a model is not totally reflected in its validation metrics. A same model could give drastically different metrics on different datasets (e.g., due to annotation quality and completeness). For instance, we identified a confounding factor: the number of lesions per scan. This variable has indeed an inverse systematic effect on the FROC, suggesting the need for its control.

Fig. 2 - FROC analysis of the model with the LUNA16 rules (challenge)

Limitations

• Strict dependence on the input normalization and spacing established by nnDetection's preprocessing plans.
• Performance may be lower on low-dose and heavily artifacted CT scans (despite some data present in the training set).
• Extremely large masses or diffuse lung diseases might not be handled as well as the more frequent solid ones.

Citation

If you use this model please cite:

Model Checkpoint:

EOLP. (2026). VAUBAN: A trained 3D Lung Nodule Detector in ONNX based on nnDetection and LIDC-IDRI. Hugging Face. https://huggingface.co/EOLP/VAUBAN

Furthermore, please cite the underlying framework and datasets used to train this model:

nnDetection Framework:

Baumgartner, M., et al. (2021). nnDetection: A Self-configuring Method for Medical Object Detection. MICCAI 2021. Lecture Notes in Computer Science, vol 12905. Springer, Cham.

LIDC-IDRI Dataset:

Armato, S.G., et al. (2011). The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans. Medical Physics, 38(2), 915-931.

Fedorov, A., et al. (2018). Standardized representation of the TCIA LIDC-IDRI annotations using DICOM. The Cancer Imaging Archive.

---

Disclaimer: This model is provided for research and assistive purposes only. It does not replace medical advice.