💻 computer science

Spatial self-supervised Peak Learning and correlation-based Evaluation of peak picking in Mass Spectrometry Imaging

This paper proposes a spatial self-supervised autoencoder-based neural network for selecting spatially structured peaks in Mass Spectrometry Imaging, accompanied by a novel evaluation framework using expert-annotated segmentation masks that demonstrates superior performance over state-of-the-art methods across diverse datasets.

Original authors: Philipp Weigand, Nikolas Ebert, Shad A. Mohammed, Denis Abu Sammour, Carsten Hopf, Oliver Wasenmüller

Published 2026-03-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Philipp Weigand, Nikolas Ebert, Shad A. Mohammed, Denis Abu Sammour, Carsten Hopf, Oliver Wasenmüller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery inside a human body. You have a high-tech camera (Mass Spectrometry Imaging, or MSI) that takes a picture of a tiny slice of tissue. But instead of seeing colors, this camera sees a massive list of chemical signals—like a library containing millions of books, where every book represents a different molecule.

The problem? Most of these "books" are just noise, static, or irrelevant junk. Only a few contain the actual clues (biomarkers) that tell you if the tissue is healthy or diseased. The process of finding these specific clues is called Peak Picking.

Here is the breakdown of the paper's solution, explained simply:

1. The Old Way: Looking at Books One by One

Previously, scientists tried to find these clues by looking at each tiny spot in the tissue individually. It was like trying to find a specific sentence in a library by reading every single book in isolation, ignoring the fact that the books are arranged on shelves in a specific order.

The Flaw: This method often picked up "noise" (random static) as if it were a clue, or missed real clues that only made sense when you looked at the neighborhood around them. It was inconsistent and hard to judge if the detective was doing a good job.

2. The New Solution: S3PL (The "Smart Neighborhood Watch")

The authors, Philipp Weigand and his team, built a new AI detective called S3PL (Spatial Self-supervised Peak Learning).

Think of S3PL not as a person reading one book at a time, but as a smart neighborhood watch.

How it works: Instead of looking at a single spot, S3PL looks at a small "patch" of the tissue and the surrounding area all at once. It uses a special neural network (a type of AI) to learn which chemical signals are "important" based on their shape and where they appear in the tissue.
The "Attention Mask": Imagine the AI puts on a pair of smart glasses. These glasses have a filter (an attention mask) that highlights the interesting chemical signals and blurs out the boring noise. It learns to do this by trying to rebuild the picture of the tissue from scratch. If it can rebuild the picture well using only the highlighted signals, it knows it picked the right clues.
The Result: It finds the "spatially structured" peaks—meaning it finds molecules that are actually organized in a pattern (like a tumor shape) rather than random specks of dust.

3. The New Scorecard: The "Map Check"

Even with a great detective, how do you know they are right? Usually, scientists had to guess or use fake data to test their methods.

The authors introduced a new way to grade the performance: The Map Check.

The Setup: Imagine a pathologist (a medical expert) has already drawn a map of the tissue, marking exactly where the tumor is and where the healthy tissue is. This is the "Ground Truth."
The Test: The AI picks its chemical clues. The researchers then check: "Do the chemical images created by these clues match the expert's map?"
The Score: They use a math tool (Pearson Correlation) to see how well the AI's chemical map lines up with the expert's drawing. If the AI picks a clue that lights up exactly where the tumor is, it gets a high score. If it picks a clue that lights up randomly, it gets a low score.
The Innovation: They didn't just use one strict rule for "good enough." They tested the AI against a range of strictness levels (from "very picky" to "somewhat lenient") to get a fair, average score. This ensures the method works even if the data is a bit messy.

4. The Results: A Better Detective

The team tested their new AI (S3PL) against the best existing methods using real patient data (brain tumors, kidney cancer, and colon cancer).

The Outcome: S3PL consistently outperformed the others. It found the right chemical clues more often and ignored the noise better.
Why it matters: In the real world, this means doctors and researchers can get clearer pictures of diseases faster. It helps in identifying biomarkers for new drugs or diagnosing cancer more accurately, without getting lost in the millions of data points.

Summary Analogy

Imagine you are trying to find the best spots to plant flowers in a giant, messy garden (the tissue).

Old methods were like walking around and picking every green leaf you saw, hoping some were flowers. You ended up with a lot of weeds.
S3PL is like a smart drone that flies over the garden, looks at the soil patterns and the arrangement of plants, and only picks the spots that actually look like flower beds.
The new evaluation is like having a master gardener's blueprint. You compare the drone's list of spots to the blueprint to see how accurate it was.

This paper gives us a smarter way to clean up the data and a better ruler to measure how well we are doing it.

1. Problem Statement

Mass Spectrometry Imaging (MSI) generates vast, complex datasets containing molecular distributions across tissue samples. A critical preprocessing step, peak picking, is required to reduce data volume by filtering distinct signals while removing noise and artifacts.

Limitations of Existing Methods: Current approaches (e.g., MALDIquant, Cardinal) typically process spectra independently, ignoring the spatial context inherent in MSI data. While some methods attempt to use spatial information (e.g., SPUTNIK, Lieb et al.), they often lack robustness across diverse datasets.
Evaluation Gap: Existing evaluation methods rely heavily on synthetic data or small, manually curated subsets of ion images. There is a lack of a consistent, quantitative evaluation framework for spatially structured peak picking (methods that select peaks based on their spatial organization) using real-world profile MSI data.

2. Methodology

The authors propose two main contributions: a novel neural network architecture for peak picking and a new evaluation framework.

A. Spatial Self-Supervised Peak Learning (S3PL)

S3PL is an end-to-end trainable neural network designed to select peaks that exhibit spatial structure.

Architecture: It utilizes a streamlined 3D Convolutional Autoencoder.
- Input: Spectral patches ( $x$ ) containing a central spectrum and its neighbors, capturing both spectral ( $m/z$ ) and spatial context.
- Attention Mechanism: The network learns a continuous attention mask via a 3D convolution. This mask is multiplied element-wise with the central spectrum to highlight informative $m/z$ values for reconstruction.
- Reconstruction: A 3D transposed convolution reconstructs the input patch. The model is trained to minimize the Mean Squared Error (MSE) between the original and reconstructed spectra.
Peak Selection Process:
1. Training: The network learns to reconstruct spectra by focusing on spatially relevant peaks.
2. Inference: The 3D convolution layer is frozen. It is applied to the entire dataset to generate an attention mask for every spectral patch.
3. Selection: The top- $z$ highest activations in the attention mask are extracted per patch. The final peak list consists of the $n$ most frequently occurring $m/z$ values across all patches.

B. Correlation-Based Evaluation Procedure

To address the lack of quantitative evaluation for spatially structured methods, the authors introduce a metric based on expert-annotated segmentation masks.

Ground Truth Definition: Instead of manually labeling individual peaks, the method uses a segmentation mask (e.g., tumor vs. healthy tissue) as the ground truth for spatial structure.
Metric Calculation:
1. Pearson Correlation Coefficient (PCC): The correlation between every ion image and the expert segmentation mask is calculated.
2. Thresholding: Ion images with a PCC $\ge T_{PCC}$ are classified as "positive" (informative/structured); others are "negative."
3. Robustness: To avoid bias from a single threshold, the authors calculate F1-scores across a range of thresholds ( $T_{PCC} \in \{0.3, 0.4, 0.5, 0.6\}$ ) and report the mean, termed mean Spatial Correlation F1-score (mSCF1).

3. Key Contributions

S3PL Network: The first deep learning-based peak picking method that explicitly leverages spatio-spectral information via self-supervised learning and attention mechanisms to select spatially structured peaks.
Novel Evaluation Framework: A quantitative, scalable evaluation procedure using expert segmentation masks and PCC-based correlation, eliminating the need for labor-intensive manual peak annotation.
Comprehensive Benchmarking: A rigorous comparison against state-of-the-art methods (MALDIquant, msiPL, Lieb et al., SPUTNIK) across diverse public datasets.

4. Results

The authors evaluated S3PL on three diverse public MSI datasets (Glioblastoma, Renal Cell Carcinoma, Colorectal Adenocarcinoma) and one dataset for qualitative analysis (Gastrointestinal Stromal Tumor).

Quantitative Performance:
- GBM Dataset: S3PL achieved the highest mSCF1 (49.6), outperforming the second-best method (msiPL) by +9.3%.
- RCC Dataset: S3PL achieved the highest mSCF1 (47.0), outperforming the second-best (MALDIquant) by +9.9%.
- CAC Dataset: S3PL achieved the highest mSCF1 (60.1), outperforming the second-best (Lieb et al.) by +11.3%.
- S3PL consistently outperformed other methods on the majority of individual tissue sections.
Qualitative Performance (GIST Dataset):
- Visual inspection showed that classical methods (MALDIquant) and single-spectrum deep learning methods (msiPL) selected many high-intensity peaks lacking spatial structure (outliers).
- S3PL successfully selected only peaks corresponding to spatially structured ion images, filtering out noise and isolated artifacts.
Ablation Studies:
- Optimal hyperparameters (spectral patch size $p$ , kernel depths $d_1/d_2$ , number of peaks $z$ ) vary by dataset, indicating the need for dataset-specific tuning.
- The method demonstrated robustness when the final peak count ( $n$ ) was determined on one tissue section and applied to others, with performance drops typically under 10%.

5. Significance

Advancement in MSI Preprocessing: S3PL bridges the gap between spectral analysis and spatial context, proving that deep learning can effectively utilize spatial priors for peak picking without requiring labeled training data (self-supervised).
Standardized Evaluation: The proposed mSCF1 metric provides a reproducible, objective standard for comparing spatially structured peak picking methods, addressing a critical bottleneck in the field where evaluations were previously qualitative or limited to synthetic data.
Scalability: The approach is computationally efficient (fewer parameters than msiPL) and adaptable to various ionization sources (MALDI, DESI) and mass analyzers (FTICR, Orbitrap, TOF).
Future Impact: The authors highlight the need for more public datasets with expert-annotated segmentation masks to further advance the field and suggest future work on automating the selection of the final peak count ( $n$ ).