Transcriptome-supervised classification of tissue morphology using deep learning

Deep learning has proven to successfully learn variations in tissue and cell morphology. Training of such models typically relies on expensive manual annotations. Here we conjecture that spatially resolved gene expression, e.i., the transcriptome, can be used as an alternative to manual annotations. In particular, we trained five convolutional neural networks with patches of different size extracted from locations defined by spatially resolved gene expression. The network is trained to classify tissue morphology related to two different genes, general tissue, as well as background, on an image of fluorescence stained nuclei in a mouse brain coronal section. Performance is evaluated on an independent tissue section from a different mouse brain, reaching an average Dice score of 0.51. Results may indicate that novel techniques for spatially resolved transcriptomics together with deep learning may provide a unique and unbiased way to find genotype-phenotype relationships.

Improving Recall of In Situ Sequencing by Self-Learned Features and a Graphical Model

Image-based sequencing of mRNA makes it possible to see where in a tissue sample a given gene is active, and thus discern large numbers of different cell types in parallel. This is crucial for gaining a better understanding of tissue development and disease such as cancer. Signals are collected over multiple staining and imaging cycles, and signal density together with noise makes signal decoding challenging. Previous approaches have led to low signal recall in efforts to maintain high sensitivity. We propose an approach where signal candidates are generously included, and true-signal probability at the cycle level is self-learned using a convolutional neural network. Signal candidates and probability predictions are thereafter fed into a graphical model searching for signal candidates across sequencing cycles. The graphical model combines intensity, probability and spatial distance to find optimal paths representing decoded signal sequences. We evaluate our approach in relation to state-of-the-art, and show that we increase recall by $27\%$ at maintained sensitivity. Furthermore, visual examination shows that most of the now correctly resolved signals were previously lost due to high signal density. Thus, the proposed approach has the potential to significantly improve further analysis of spatial statistics in in situ sequencing experiments.