Automated Tissue Classification and Candidate Biomarker Feature Extraction in Mass Spectrometry Imaging Based on Interpretable Deep Learning Algorithm.

RATIONALE: Mass spectrometry imaging (MSI) generates high-dimensional spatial-spectral data that requires efficient computational methods for tissue classification and candidate biomarker feature extraction. Deep learning offers a promising approach, yet the interpretability of model predictions and identification of biologically relevant spectral features remain challenging.

METHODS: A comprehensive computational pipeline was developed for automatic tissue layer classification of a public mouse urinary bladder MSI dataset. Building upon prior work that compared manual tissue layer labels with those automatically generated via spectral preprocessing, t-SNE, and hierarchical clustering, in this study we separately use each type of class label to train convolutional neural networks (CNNs) for supervised classification. Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) were employed to compute layer-specific summed importance scores to evaluate each mass spectral feature and extract class-discriminative features.

RESULTS: On the mouse urinary bladder MSI dataset, both manual labels and cluster-derived labels (t-SNE + hierarchical clustering) enabled the CNN model to achieve training accuracies exceeding 0.9 for classifying three tissue layers. Interpretability methods successfully identified discriminative m/z features, including known lipids such as SM(34:1) (m/z 741.54) and PC(34:1) (m/z 798.54), consistent with previously reported biological markers. Compared to the intensity values, the importance scores of the top class-discriminative features generated by both interpretability methods exhibited a sharper contrast and superior ability to delineate tissue-layer-specific distributions in their ion images. Furthermore, evaluation on an independent colorectal cancer dataset yielded a test accuracy of 0.758, suggesting that cross-patient generalizability varied across different data sources.

CONCLUSIONS: This study presented an effective deep learning framework for accurate and interpretable tissue classification in MSI data. The CNN modeling and deep learning interpretability provided a robust approach for both automated segmentation and biological discovery, facilitating the identification of spatially resolved metabolic features in tissue sections.