Multi-center evaluation of tissue classification by mass spectrometry imaging

Background It is possible perform tissue phenotyping based on mass spectrometric imaging data. However, comprehensive studies assessing the variation across different sites and their impact on tissue classification are largely lacking. Here, we have evaluated how well tissue classification based on Matrix-Assisted-Laser-Desorption/Ionization (MALDI) mass spectrometric imaging (MSI) can generalize across sites. Design A tissue microarray (TMA) human FFPE samples representing 6 different tumors entities (leiomyoma, seminoma, mantle cell lymphoma, melanoma, invasive ductal carcinoma of the breast and squamous cell carcinoma of the lung) was used. These samples were prepared for MALDI-MSI and measured at three different sites using a standard protocol. The baseline performance of the classification task was evaluated by using cross-validation on the individual TMAs. We then evaluated how well a classification model based on data from two sites performed on the data of the remaining site. Results Selecting the relevant mass features for training the classifier was necessary for performance of classification. Treating the entire cores as representative for the respective tumor entity, the baseline performance for the classification was an accuracy of 82.6% correct classifications. Accuracy over sites was 74.1%. With detailed histological annotations the classification accuracy was 92% on the individual TMAs and to 84% accuracy when applied over sites. Conclusion Initial results indicate that MALDI-MSI can be performed at a level that allows relevant multi-center research studies. MALDI imaging based tissue classifiers are able to generalize across sites. A detailed histological annotation of the tissue improves the performance of the classifiers.