IonMorphNet: Generalizable Learning of Ion Image Morphologies for Peak Picking in Mass Spectrometry Imaging

Official repository of the CVPR 2026 Workshop paper "IonMorphNet: Generalizable Learning of Ion Image Morphologies for Peak Picking in Mass Spectrometry Imaging".

Philipp Weigand*, Niels Nawrot*, Nikolas Ebert, Carsten Hopf & Oliver Wasenmüller | *Equal Contribution
CeMOS - Research and Transfer Center, University of Applied Sciences Mannheim

 

Installation

1. Clone this repository

git clone https://github.com/CeMOS-IS/IonMorphNet.git
cd IonMorphNet

2. Create and activate a conda environment

conda create -n ionmorphnet python=3.9.23
conda activate ionmorphnet

3. Install dependencies and the package

pip install -r requirements.txt
pip install -e .

Datasets

The package expects MSI datasets in the following structure:

IonMorphNet
└── data
    └── datasets
        ├── <dataset-id-1>
        │   ├── <dataset-id-1>.imzML
        │   └── <dataset-id-1>.ibd
        ├── <dataset-id-2>
        │   ├── <dataset-id-2>.imzML
        │   └── <dataset-id-2>.ibd
        └── ...

We provide our created morphology labels for each dataset at this path:

data/labeling/csv/<dataset-id>.csv

You can find the corresponding dataset under

https://metaspace2020.org/dataset/<dataset-id>

Training

Run the following command and specify the validation and test split with --val_files and --test_files. All remaining files will be used for training. The following split was used for our trainings.

python -m msianalyzer.training.train_msi_classifier \
  --timm_model "convnextv2_tiny" \
  --val_files "2017-09-25_19h48m56s.imzML,2023-07-04_10h26m22s.imzML,2025-05-20_08h21m56s.imzML,2025-06-18_03h24m09s.imzML,2025-09-24_18h51m29s.imzML" \
  --test_files "2016-10-01_12h21m40s.imzML,2016-10-01_12h27m29s.imzML,2016-10-25_14h30m16s.imzML,2017-03-10_19h59m17s.imzML,2017-03-17_17h20m49s.imzML"

Evaluation

Make sure to provide the mSCF1 evaluation datasets (GBM, CAC) with corresponding segmentation masks in the following folder structure:

├── IonMorphNet
    ├── data
        ├── mSCF1_Evaluation_Datasets
            ├── dataset1
            │   ├── masks
            │   │   ├── file1_mask.npy
            │   │   ├── file2_mask.npy
            │   │   └── ...
            │   ├── file1.imzML
            │   ├── file1.ibd
            │   ├── file2.imzML
            │   ├── file2.ibd
            │   └── ...
            ├── cac
            │   ├── masks
            │   │   ├── 40TopL_mask.npy
            │   │   ├── 40TopL_mask.npy
            │   │   └── ...
            │   ├── 40TopL.imzML
            │   ├── 40TopL.ibd
            │   └── ...
            └── ...

To use our trained ConvNeXt-V2-tiny model, download it from here and extract the zip in the IonMorphNet/data/models/ directory. Then run the Peak Picking evaluation:

python -m msianalyzer.evaluate.evaluate_mSCF1_peak_quality \
--run-dir 20260204-164840_convnextv2_tiny \
--informative-classes structured,negative,localized

where --run-dir corresponds to the specific model foldername in IonMorphNet/data/models/ that should be used for evaluation. The results will be stored within that folder in /evaluation_mSCF1/<date-time>/evaluation_results.csv.

Application: Classification of all ion images in imzML/ibd file

python -m msianalyzer.evaluate.classify_ion_images \
--run-dir <model_folder_name> \
--imzml-folderpath "/path/to/folder/with/imzML/files"

where --run-dir corresponds to the model foldername in IonMorphNet/data/models/ that should be used. The results will be stored within that folder in /morphology_predictions. --imzml-folderpath corresponds to the folder that contains the imzML file(s) that should be anaylzed.

Troubleshooting

No datasets are found. Verify that:

• dataset files are located at IonMorphNet/data/datasets/<dataset-id>/
• each dataset directory contains two files:
  • <dataset-id>.imzML
  • <dataset-id>.ibd
• the file extension is exactly .imzML with the correct capitalization
• each dataset has a matching CSV file in IonMorphNet/data/labeling/csv/
• the CSV filename exactly matches the dataset name

Citing

If you use this code in your research, please cite our paper.

@inproceedings{weigand2026ionmorphnet,
  title={IonMorphNet: Generalizable Learning of Ion Image Morphologies for Peak Picking in Mass Spectrometry Imaging},
  author={Weigand, Philipp and Nawrot, Niels and Ebert, Nikolas and Hopf, Carsten and Wasenm{\"u}ller, Oliver},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year={2026}
}