Deep-Learning Based Cell Segmentation and Deconvolution in Spatial Transcriptomics

Mena Soliman Asaad Kamel, Amrut Sarangi, Cindy Qin, Het Barot, Pavel Senin, Sergio Villordo, Sunaal Mathew, Albert Pla Planas, Ziv Bar-Joseph

Abstract: Next-generation sequencing (NGS) technologies made it possible to study the cell structure and composition of a given tissue while preserving spatial context, giving birth to the field of spatial transcriptomics. However, NGS technologies such as Visium provide gene counts at spot locations that contain up to 50 cells which limits researchers from studying the tissue at a single-cell level. Computational techniques such as Cell2Location can deconvolve the gene counts to obtain the proportions of cell types at each spot. Those approaches, however, do not provide a way to assign the predicted cell types to the individual cells within a Visium spot. In parallel, the digital pathology field has seen rapid growth with the introduction of deep convolutional neural networks that can segment and classify cells with high accuracies, matching the level of the pathologist.The work presented herein introduces a solution that combines recent cell deconvolution algorithms and deep learning-based cell segmentation models to estimate individual cell type assignments within a Visium spot. Using the inferred cell types and locations, spatial statistical tests can be run to quantify the proximity of target cell types to tumor boundaries. This provides the basis for identifying the degree of infiltration of immune cells within tumors. The process starts by deconvolving each Visium spot using Cell2Location. Separately, cells in the high-resolution H&E tissue image are segmented using cell segmentation convolutional neural networks such as CellPose and HoverNet which are fine-tuned for optimal performance on the target tissue and species. For each Visium spot, the underlying cells are counted and clustered based on morphological features. By combining the clustering results with the cell proportions obtained by Cell2Location, one can obtain preliminary cell type assignments which allows for studying the tissue landscape at a pseudo single-cell level. This unlocks the ability to study the spatial organization of cells within the tissue, providing a way to automatically identify patterns of immune cell infiltration in human lung cancer biopsies.Validation involves gathering pathologist annotations for regions of tumors and immune cell aggregates within Visium Experiments. For each annotated region, the proportions of Tumor-infiltrating lymphocytes (TILs) within the tumor are computed as a function of distance from the tumor boundary. This is compared to the proportion of TILs within a radial distance from the tumor boundary using a two-sided Z-test. In a pool of 12 human lung cancer biopsies and their corresponding Visium experiments, preliminary results show that the cell types predicted by cell deconvolution and the degree of infiltration agree with the annotations provided by the pathologist. This suggests the feasibility of the proposed approach at automatically identifying TILs. Further studies in this line may open the possibility of quantifying the impact of the distribution of immune structures on tumor progression and anti-tumor response.

External IDs:doi:10.1145/3584371.3613050