PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Tumor Classification in Imaging Mass Spectrometry

Jens Behrmann, Christian Etmann, Tobias Boskamp, Rita Casadonte, Jörg Kriegsmann, Peter Maass

arXiv:1705.01015v3stat.MLcs.LG

TL;DR

Tumor typing from IMS requires analysis of complex, high-dimensional spectra, while conventional workflows separate feature extraction and classification. This paper adapts CNNs to spectra and analyzes their predictions with sensitivity analysis; cross-validation reports competitive performance on two tasks, with IsotopeNet outperforming comparison methods and reaching 84.5% spot-level balanced accuracy on ADSQ.

  • Problem

    IMS tumor classification requires robust extraction of useful information from high-dimensional spectra, while common workflows separate feature extraction and classification.

    Source: §I · pp. 1–2

  • Method

    The paper adapts convolutional neural networks to mass spectra and uses sensitivity analysis to interpret the relationship between input m/z values and predicted class probabilities.

    Source: C. Architectures for IMS · p. 4, D. Interpretation via Sensitivity… · p. 5

  • Results

    IsotopeNet performed better than ResidualNet and ROC/LDA, with spot-level balanced accuracy of 84.5% for task ADSQ.

    Source: B. Model comparison · p. 6

  • Takeaways & Limitations

    The authors present the study as a starting point for further development of end-to-end learning methods for IMS classification.

    Source: A. Summary · p. 8

    The study reports frequent gaps between training and test performance due to overfitting, highlighting the need for better regularization on small-sample tasks.

    Source: B. Future work · p. 8

Abstract

from arXiv · show

Motivation: Tumor classification using Imaging Mass Spectrometry (IMS) data has a high potential for future applications in pathology. Due to the complexity and size of the data, automated feature extraction and classification steps are required to fully process the data. Deep learning offers an approach to learn feature extraction and classification combined in a single model. Commonly these steps are handled separately in IMS data analysis, hence deep learning offers an alternative strategy worthwhile to explore. Results: Methodologically, we propose an adapted architecture based on deep convolutional networks to handle the characteristics of mass spectrometry data, as well as a strategy to interpret the learned model in the spectral domain based on a sensitivity analysis. The proposed methods are evaluated on two challenging tumor classification tasks and compared to a baseline approach. Competitiveness of the proposed methods are shown on both tasks by studying the performance via cross-validation. Moreover, the learned models are analyzed by the proposed sensitivity analysis revealing biologically plausible effects as well as confounding factors of the considered task. Thus, this study may serve as a starting point for further development of deep learning approaches in IMS classification tasks.

I. INTRODUCTION

Imaging mass spectrometry produces spatially indexed spectra, but extracting robust features from these high-dimensional data is challenging. The paper explores supervised end-to-end CNNs adapted to spectra for tumor classification across two TMA-based tasks.

I. INTRODUCTION

  • IMS records mass spectra at spatial tissue positions, enabling molecular analysis that may support tumor subtyping and identification of primary tumors.MALDI IMS can measure multiple tumor cores arranged in a tissue microarray in a single run.Source: §I · p. 1
  • Robust feature extraction is needed to analyze high-dimensional spectra, while common workflows separate feature extraction from supervised classification.Peak detection and molecular-signature methods are among the existing feature-extraction approaches.Source: §I · pp. 1–2
  • The paper introduces fully supervised deep learning for large-scale IMS tumor classification, adapting CNN architectures to mass spectra and examining model interpretability.Earlier IMS neural-network work cited here focused on unsupervised or dimension-reduction approaches.Source: §I · p. 2
  • The study evaluates adenocarcinoma versus squamous cell carcinoma and lung versus pancreas classification using datasets previously studied by M. Kriegsmann et al. (2016) and Boskamp et al. (2016).The two tasks use eight and twelve TMAs, respectively.Source: §I · p. 2

II. METHODS

The methods describe FFPE tissue preparation and MALDI-IMS acquisition, then train CNN classifiers on individual spectra. The network design treats spectra as one-dimensional structured inputs and adapts convolutional layers to their characteristics.

A. Samples, MALDI-IMS, preprocessing

  • FFPE cores were assembled into 12 TMAs, prepared with tryptic digestion and MALDI matrix, then measured from 500–4500 m/z at 150 µm spacing using 1600 laser shots per position.Data received baseline correction after acquisition.Source: A. Samples, MALDI-IMS, preprocessing · p. 2

B. Deep Convolutional Neural Networks

  • After preprocessing, each IMS spot’s spectrum is handled separately as a data point on a predefined m/z-bin grid, analogous to a one-dimensional image.The grid need not be equidistant, unlike an image pixel grid.Source: B. Deep Convolutional Neural Networks · pp. 2–3
  • The network maps labeled spectra to class probabilities and is trained by minimizing average negative log-likelihood with weight decay, using mini-batch stochastic gradient updates and Adam.Weight decay is intended to prevent overfitting; mini-batches permit multiple parameter updates at roughly the cost of one full-set gradient step.Source: B. Deep Convolutional Neural Networks · p. 3
  • Convolutional layers combine feature maps through kernels, while residual connections let inputs bypass layers and support deep networks.When dimensions change or strided convolutions are used, the residual input is adjusted by a convolutional layer.Source: B. Deep Convolutional Neural Networks · pp. 3–4
  • IsotopeNet uses consecutive, partly strided convolutions to expand receptive fields, followed by locally connected processing that can treat neighboring spectral regions differently.The receptive fields reach the size of an isotopic pattern after the locally connected layer.Source: Figure 2 · p. 4, B. Deep Convolutional Neural Networks · p. 4

C. Architectures for IMS

The architecture adapts convolutional networks to the structure of IMS spectra by limiting local grouping to roughly the size of large isotope patterns. A locally connected layer then processes nearby features with region-specific weights.

C. Architectures for IMS

  • CNNs group neighboring m/z-bins, while pooling or strided convolution progressively combines them into lower-dimensional representations.This applies convolutional feature extraction to spectra treated as structured data on an m/z grid.Source: C. Architectures for IMS · p. 4
  • Because isotope patterns are local but protein-related peptide patterns may span irregular mass ranges, the architecture restricts local grouping to roughly the size of large isotope patterns.The design assumes neighboring m/z-bins are correlated, particularly where a measured peak spreads across bins.Source: C. Architectures for IMS · p. 4
  • After two subsampling operations, IsotopeNet reduces spectra from 27,268 dimensions to 1,820 variables, then processes neighboring variables with a locally connected layer.Unshared weights let the layer handle each local region differently, allowing focus on task-relevant peptides.Source: C. Architectures for IMS · p. 4
  • The locally connected layer complements convolutional processing by allowing different local regions to use different weights.The paper contrasts this region-specific processing with the shared convolutional transforms.Source: C. Architectures for IMS · p. 4

D. Interpretation via Sensitivity Analysis

The paper interprets model behavior by measuring how input m/z values relate to predicted class probabilities. It scales sensitivities by input variation and averages them over training samples for broader model-level interpretation.

D. Interpretation via Sensitivity Analysis

  • Sensitivity analysis uses the gradient of a class probability with respect to a spectrum to measure prediction sensitivity to changes in input m/z-values.The sensitivity can be computed by backpropagation because the network is differentiable with respect to its input.Source: D. Interpretation via Sensitivity… · p. 5
  • Sensitivities are scaled per m/z dimension by the corresponding input standard deviation to make sensitivities across dimensions comparable.The passage identifies input-scale dependence as the reason for this normalization.Source: D. Interpretation via Sensitivity… · p. 5
  • Because a gradient is computed for each sample, the paper averages sensitivities over the training set to make more general statements about model behavior.A single-sample gradient supports interpretation by example only.Source: D. Interpretation via Sensitivity… · p. 5

A. Datasets and Evaluation

The study evaluates tumor classification on two MALDI-IMS tasks using randomized four-fold cross-validation at the tissue-microarray level. It compares deep networks with a discriminative-m/z feature-selection and LDA baseline.

A. Datasets and Evaluation

  • The two tasks are adenocarcinoma versus squamous cell carcinoma (ADSQ), using 4,672 ROI spectra, and lung versus pancreas tumor typing (LP), using 27,475 spectra from entire cores.LP includes spots with non-tumor cells, whereas ADSQ uses spectra within pathologist-marked regions of interest.Source: A. Datasets and Evaluation · p. 5
  • Performance is evaluated by randomized four-fold cross-validation on TMA level, with balanced accuracy measured at spot level and by majority-vote aggregation at core level.The study uses the median balanced accuracy across the four cross-validation runs.Source: A. Datasets and Evaluation · p. 5
  • The ROC/LDA baseline ranks m/z-values by their Mann–Whitney–Wilcoxon statistic, selects 5–100 features, and classifies them with linear discriminant analysis.The feature count varies to provide a range of baseline results.Source: A. Datasets and Evaluation · p. 5

B. Model comparison

Across the two tumor-classification tasks, IsotopeNet outperforms ResidualNet and ROC/LDA in the reported comparisons. Its fold-level performance varies, and the authors caution that small differences may be premature to interpret.

B. Model comparison

  • 84.5% spot-level balanced accuracy for task ADSQ; IsotopeNet outperforms ResidualNet and ROC/LDA overall.Figure 3 reports the four-fold comparison; its table gives medians and interquartile ranges for the neural networks and best/worst results across feature counts for ROC/LDA.Source: Figure 3 · p. 6, B. Model comparison · p. 6
  • IsotopeNet has a higher median balanced accuracy than the other methods on both tasks in the spot-level fold boxplots.The boxplots show variation across folds, including outliers for both neural-network methods.Source: B. Model comparison · p. 6, Figure 3 · p. 6
  • Task LP appears easier for all methods, which the authors relate to more spectra being available and the likely ease of distinguishing primary tumors.The discussion notes that greater data availability is especially relevant for deep learning.Source: B. Model comparison · p. 6
  • Fold-to-fold variance remains substantial, so the authors caution that conclusions based on small performance differences may be premature.They also state that more extensive parameter searches could improve results, especially for ResidualNet, but are outside this study's scope.Source: B. Model comparison · p. 6
  • On task ADSQ, IsotopeNet reaches about 95% training balanced accuracy, while ResidualNet exceeds 98%.The authors suggest ResidualNet's larger parameter count may explain this difference.Source: B. Model comparison · p. 6

C. Interpretation of models

The sensitivity analysis examines the selected IsotopeNet’s output–input relationship in the spectral domain. It links biologically plausible signals to tumor markers in ADSQ while exposing measurement-related confounding in LP.

Interpretation of models

  • The section frames biological interpretation as necessary for assessing model strengths and weaknesses; the baseline’s discriminative m/z-values can be followed by MS/MS identification.For ADSQ, prior work identified differentially expressed peptide ions using MS/MS on tissue digest (M. Kriegsmann et al., 2016).Source: C. Interpretation of models · p. 7
  • The analysis averages class-specific sensitivity across training spectra from the best cross-validation fold of the best of four consecutive runs.The resulting sensitivity has the dimensions of a raw spectrum, allowing interpretation in the input domain.Source: C. Interpretation of models · p. 7
  • In ADSQ, sensitivity peaks at 1406.6 Da and 1410.7 Da align with CK7 and CK5 peptides; other peaks align with CK15 and HSP27 peptides.CK5 and CK7 are established IHC markers, while CK15 and HSP27 are described as potential new markers.Source: C. Interpretation of models · p. 8
  • For LP, oscillations across 2100–2900 Da are biologically implausible and may reflect measurement artifacts that confound lung-versus-pancreas classification.The tissues were measured on separate TMAs, and this confounding may explain the task’s higher reported performance.Source: C. Interpretation of models · p. 8

A. Summary

The paper presents deep neural networks for IMS tumor classification, reports competitive performance, and proposes sensitivity analysis to interpret model behavior and assess its validity.

Summary

  • The proposed deep-learning approach was tested on challenging real-world IMS tumor datasets and showed competitive classification performance.The adapted architecture outperformed a standard deep-learning approach.Source: A. Summary · p. 8
  • Sensitivity analysis linked the model’s ADSQ discrimination to known biomarkers and revealed artifacts affecting lung-versus-pancreas discrimination.Separate measurements can support classification through measurement-characteristic differences.Source: A. Summary · p. 8
  • The study presents its work as a starting point for further end-to-end learning research in IMS.The authors note that convolutional-network training requires a powerful GPU, while batch training allows efficient processing of large sample sets.Source: A. Summary · p. 8

B. Future work

Future work targets domain-informed models, better regularization for small datasets, and methods that address biological and technical variation in IMS.

Future work

  • The authors plan to incorporate domain knowledge, including peak shapes and known biomarkers, into the network.They propose using biomarkers described in the paper to guide network behavior.Source: B. Future work · p. 8
  • Frequent gaps between training and test performance indicate overfitting; the authors identify stronger regularization as important for small-sample tasks.They suggest methods beyond the applied weight decay and dropout, including constraining the model to fewer peaks and checking this with sensitivity analysis.Source: B. Future work · p. 8
  • Future work will address high biological and technical variation and examine technical errors and patient-level variation.Data augmentation is also proposed as a possible way to improve robustness; evaluation on other large, challenging classification tasks is needed to better understand model strengths and weaknesses.Source: B. Future work · p. 9

FUNDING

The authors acknowledge support from German federal and research-funding bodies.

Funding

  • Funding came from the German Federal Ministry of Education and Research and the German Science Foundation.The federal support was through the “KMU-innovativ: Medizintechnik” Programme; the foundation supported RTG 2224 “π3: Parameter Identification - Analysis, Algorithms, Applications”.Source: FUNDING · p. 9
Loading 1705.01015v3…