MouseLungLesionSeg v1.0: A Two-Staged AI Approach for Lung Lesion Segmentation on MRI for Mice Tumor Models
Year: 2026
Type: dataset
Abstract: Description This record provides trained nnU-Net v2 model weights for automated segmentation of mouse lungs and lung lesions in T2-weighted (T2W) magnetic resonance images. It also includes representative, de-identified sample MRI volumes with corresponding reference ground-truth annotations and AI-generated predictions for demonstrating and evaluating the inference workflow. The pipeline contains two independently trained 3D full-resolution nnU-Net models: A lung-segmentation model that identifies the mouse lung region. A lung-lesion segmentation model that identifies suspected lesions from the same input MRI. The two predictions are subsequently merged. Predicted lesion voxels are retained only when they occur within the predicted lung mask. The resulting combined segmentation uses the following label convention: 0: Background 1: Lung without predicted lesion 2: Predicted lesion within the predicted lung Each input is a single-channel T2W MRI volume in compressed NIfTI format (.nii.gz) and follows the nnU-Net naming convention CASE_IDENTIFIER_0000.nii.gz. The model archive includes the final trained checkpoints and the nnU-Net metadata required for inference, including dataset.json, dataset_fingerprint.json, and plans.json. The sample-data package contains, as applicable: T2W mouse MRI volumes Reference ground-truth lung and lesion annotations AI-generated lung predictions AI-generated lesion predictions Combined three-label AI segmentations Visual comparisons of the reference and AI annotations Ground-truth annotations and AI-generated annotations are stored separately and clearly identified. The AI outputs are provided to demonstrate model behavior and reproducibility; they must not be interpreted as additional ground truth. The associated GitHub repository provides installation instructions, inference scripts, prediction-merging utilities, and tumor-burden analysis tools: GitHub Model provenance Framework: nnU-Net v2 Configuration: 3d_fullres Trainer/plans: nnUNetTrainer__nnUNetPlans__3d_fullres Training fold: all Checkpoint: checkpoint_final.pth Lung model dataset identifier: Dataset101_mice_final Lung-lesion model dataset identifier: Dataset101_MRI-Lung Recommended Python version: Python 3.10 or newer Recommended inference device: CUDA-compatible GPU; CPU inference is also supported The original lung-model metadata names input channel 0 as CT and specifies CTNormalization. These original training metadata are retained because nnU-Net uses the saved plans during inference. However, the public pipeline input is a single T2-weighted MRI volume named according to the *_0000.nii.gz convention. Research-use disclaimer This software and the accompanying models are provided for research use only. They are not medical devices and have not been validated or approved for clinical diagnosis, treatment planning, or patient-care decisions. If you use these models, please cite this Zenodo record, the associated GitHub repository, the related publication [COMING SOON], and the nnU-Net framework.
Institutions: European University of Bangladesh, National Cancer Institute
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204251710
Cites:
Cited by:
Related to: