IonMorphNet: Generalizable Learning of Ion Image Morphologies for Peak Picking in Mass Spectrometry Imaging
Philipp Weigand, Niels Nawrot, Nikolas Ebert, Carsten Hopf, Oliver Wasenmüller · 2026
Read IonMorphNet: Generalizable Learning of Ion Image Morphologies for Peak Picking in Mass Spectrometry Imaging by Philipp Weigand, Niels Nawrot, Nikolas Ebert, Carsten Hopf, Oliver Wasenmüller (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-19369.txt.
Introduction
Weigand and colleagues introduce IonMorphNet, a model for peak picking in mass spectrometry imaging that learns ion image shapes without task-specific supervision. Peak picking is a basic preprocessing step, with each sample holding hundreds to thousands of ion images. Existing methods need careful per-dataset tuning and often fail across acquisition protocols. IonMorphNet is trained as a spatial-structure-aware representation model on 53 public datasets, using six structural classes of ion image patterns, and it then assesses new ion images with no further tuning.
Catalog record
| CREATOR | Philipp Weigand, Niels Nawrot, Nikolas Ebert, Carsten Hopf, Oliver Wasenmüller |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33269 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.19369
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
IonMorphNet: Generalizable Learning of Ion Image Morphologies for Peak Picking in Mass Spectrometry Imaging Philipp Weigand, Niels Nawrot, Nikolas Ebert, Carsten Hopf, Oliver Wasenmüller Abstract Peak picking is a fundamental preprocessing step in Mass Spectrometry Imaging (MSI), where each sample is represented by hundreds to thousands of ion images. Existing approaches require careful dataset-specific hyperparameter tuning, and often fail to generalize across acquisition protocols. We introduce IonMorphNet, a spatial-structure–aware representation model for ion images that enables fully data-driven peak picking without any task-specific supervision. We curate 53 publicly available MSI datasets and define six structural classes capturing representative spatial patterns in ion images to train standard image backbones for structural pattern classification. Once trained, IonMorphNet can assess ion images and perform peak picking without additional hyperparameter tuning. Using a ConvNeXt V2-Tiny backbone, our approach improves peak picking performance by + 7 % +7\% mSCF1 compared to state-of-the-art methods across multiple datasets. Beyond peak picking, we demonstrate that spatially informed channel reduction enables a 3D CNN for patch-based tumor classification in MSI. This approach matches or exceeds pixel-wise spectral classifiers by up to + 7.3 % +7.3\% Balanced Accuracy on three tumor classification tasks, indicating meaningful ion image selection. The source code and model weights are available at https://github.com/CeMOS-IS/IonMorphNet . 1 Introduction Figure 1 : Overview of IonMorphNet, a foundational ion image encoder for Mass Spectrometry Imaging (MSI). IonMorphNet is pre-trained on morphology classification across diverse organisms, organs and instrument types. The frozen encoder performs ion image assessment, thereby enabling spatially grounded peak picking across diverse datasets. Furthermore, spatially grounded peak picking enables tumor classification with 3D CNNs which can improve Balanced Accuracy by up to +7.3\% compared to pixelwise approaches. Mass Spectrometry Imaging (MSI) enables label-free analysis of molecular spatial distributions in biological tissues which is important for pharmaceutical research, drug development 35 , discovery of candidate biomarkers 6 and histopathological tissue analysis 34 . MSI produces biomolecular images with channel dimensions of up to tens of thousands of channels (m/z values). Untargeted analysis requires prefiltering of distinct signals and removal of artifacts and noise to enhance the information content in MSI data and the single channel images called ion images . The standard filtering strategy is peak picking 17 ; 9 , which is currently challenged by rule-based spatially informed approaches 26 ; 20 ; 4 ; 3 as well as deep-learning-based methods 1 ; 41 . However, all existing methods need hyperparameter tuning, and in the case of deep-learning-based approaches, training from scratch is necessary for each MSI dataset. This has the potential to yield favourable outcomes on particular datasets for which a method was developed. However, it can result in suboptimal generalization across other datasets, as demonstrated recently 41 . While powerful image encoders dominate the domain of natural images 14 ; 33 ; 36 , they have so far been absent from the MSI domain. Tasks like classification, segmentation, and object detection have not been widely applied in MSI due to the limited availability and high cost of annotated samples. Furthermore, the tissue sections originate from different organisms and organs, that are measured on different devices and require expert annotation to obtain labels such as tumorous and healthy tissue regions. These diverse data sources pose a challenge regarding the generalization across organisms, organs and measurement devices. Currently, most supervised tasks are performed on single datasets with their specific characteristics 7 ; 21 ; 2 ; 24 ; 22 . Therefore, the MSI domain would greatly benefit from a generalizable ion image encoder which can be applied across various types of datasets. In this work, we propose IonMorphNet, the first general-purpose image encoder for MSI ion images. It is designed to enhance spatially informed peak picking without any required hyperparameter tuning during inference. To enable cross-dataset generalization, we collected 53 publicly available diverse MSI datasets covering different organisms and organs, measured with various devices, and labeled their ion images with structure-based class labels that can be transferred to any MSI dataset. This unique dataset collection enables, for the first time, large scale training of common image encoders for structure assessment of ion images. Our IonMorphNet outperforms state-of-the-art peak picking methods by +7\% Avg. mSCF1 while requiring neither retraining nor hyperparameter tuning for new datasets. We perform ablation studies on the selected classes for peak picking, on different backbones and analyze the impact of ImageNet 12 pretraining. Furthermore, we showcase the applicability of our IonMorphNet as a preprocessing step before tumor classification. By reducing the channel dimensionality of MSI datasets, we enable the application of 3D CNNs, which exceed pixel-wise classification by up to +7.3\% Balanced Accuracy or achieve similar results. 2 Related Work 2.1 Peak Picking Classical Peak Picking. In an MSI dataset, spectra are typically represented in either profile or centroided format. Profile spectra retain the full peak shapes and typically contain the same number of m/z values in each pixel. Centroiding, or peak picking, reduces these peak shapes to a single m/z value per peak. Most peak picking methods process spectra separately from each other, without spatial context. MALDIquant 17 and Cardinal 9 , for instance, offer modular MSI analysis pipelines that encompass peak picking functionality. Such methods typically rely on signal-to-noise ratio thresholding applied to individual spectra to differentiate true peaks from background noise. However, they completely neglect the spatial information inherent in MSI data. Sparse frame multipliers 26 were proposed as the first technique to exploit both spatial and spectral (spatio-spectral) information for peak picking of structured ion images in matrix-assisted laser desorption/ionization (MALDI) time-of-flight (TOF) MSI datasets. Unlike the methods described above, another line of work operates exclusively on ion images, disregarding the spectral dimension of MSI data. The measure of spatial chaos 4 was introduced to quantify the spatial structure of peak picked ion images, while an adaptation of the Gray Level Co-Occurrence (GCO) matrix 16 was later proposed for analogous purposes 43 . A subsequent refinement of the spatial chaos measure was presented as part of a broader pipeline for identifying biologically meaningful molecular signals 30 . It is important to note that both the spatial chaos measure and the GCO matrix approach have been specifically designed to evaluate the spatial quality of pre-selected, already peak-picked ion images, and neither constitutes a peak picking method in its own right. SPUTNIK 20 takes a different approach, even though it similarly focuses on ion images, but is explicitly designed as a peak picking pipeline for profile MSI data. The method first constructs a dataset-level reference image via principal component analysis 32 , then identifies relevant peaks by correlating individual ion images against this reference. Figure 2 : Overview of all 53 collected public datasets regarding ion source, analyzer, organism and organism part. Deep-Learning-based Peak Picking. The latest peak picking approaches are both self-supervised autoencoders. msiPL 1 utilizes a variational autoencoder 23 of feed-forward-layers to reconstruct single mass spectra of a MSI dataset separately. Subsequently, relevant peaks are selected by identifying the most prominent input weights of the model and the corresponding m/z-values. S 3 PL 41 extends this approach by processing spatio-spectral patches in a self-supervised manner using 3D Convolutions and an attention mask as the bottleneck. After training, the attention mask