Multicenter Evaluation of Tissue Classification by Matrix-Assisted Laser Desorption/Ionization Mass Spectrometry Imaging.

Abstract

Many studies have demonstrated that tissue phenotyping (tissue typing) based on mass spectrometric imaging data is possible; however, comprehensive studies assessing variation and classifier transferability are largely lacking. This study evaluated the generalization of tissue classification based on Matrix Assisted Laser Desorption/Ionization (MALDI) mass spectrometric imaging (MSI) across measurements performed at different sites. Sections of a tissue microarray (TMA) consisting of different formalin-fixed and paraffin-embedded (FFPE) human tissue samples from different tumor entities (leiomyoma, seminoma, mantle cell lymphoma, melanoma, breast cancer, and squamous cell carcinoma of the lung) were prepared and measured by MALDI-MSI at different sites using a standard protocol (SOP). Technical variation was deliberately introduced on two separate measurements via a different sample preparation protocol and a MALDI Time of Flight mass spectrometer that was not tuned to optimal performance. Using standard data preprocessing, a classification accuracy of 91.4% per pixel was achieved for intrasite classifications. When applying a leave-one-site-out cross-validation strategy, accuracy per pixel over sites was 78.6% for the SOP-compliant data sets and as low as 36.1% for the mistuned instrument data set. Data preprocessing designed to remove technical variation while retaining biological information substantially increased classification accuracy for all data sets with SOP-compliant data sets improved to 94.3%. In particular, classification accuracy of the mistuned instrument data set improved to 81.3% and from 67.0% to 87.8% per pixel for the non-SOP-compliant data set. We demonstrate that MALDI-MSI-based tissue classification is possible across sites when applying histological annotation and an optimized data preprocessing pipeline to improve generalization of classifications over technical variation and increasing overall robustness.