Automatic Tissue Detection for Mass Spectrometry Imaging

Mass spectrometry imaging is a powerful technique which maps the spatial distribution of thousands of biomolecules across tissue sections. The clear delineation of tissue is an important preceding analysis step typically requiring manual intervention. We present an end-to-end method for the automatic detection of tissue in mass spectrometry images (MSIs) using same-tissue-section pairs MSIs and histological images. First, the histological tissue masks were annotated using QuPath. Second, manually acquired landmarks were used to fit to affine transforms and map the tissue masks into the MSI space. Third, we proposed metabolite-independent representations of MSIs─based on total-ion-current, root-mean-square and Shannon-entropy images─to fit a tissue-detection model. Finally, a convolutional neural network was trained to detect tissue using cross-validation in a set of 68 images featuring a variety of tissue types, organisms and spatial resolutions. Our model achieved a cross-validation accuracy, precision, recall, and Sørensen-Dice coefficient of 0.953 ± 0.030, 0.939 ± 0.047, 0.923 ± 0.056, and 0.930 ± 0.041, respectively. Using unseen test data from two different studies, our model obtained an accuracy, precision, recall, and Sørensen-Dice coefficient of 0.945 ± 0.007, 0.965 ± 0.009, 0.915 ± 0.027, and 0.935 ± 0.011, respectively.