Oral graft versus host disease : salivary gland pathology and tissue modelling

Chronic Graft versus Host Disease (cGVHD) is a consequence of dysregulated immune reconstitution after allogenic hematopoietic cell transplantation (alloHCT), a treatment for high-risk blood malignancies or immune disorders. cGVHD is a donor T-cell-driven rejection of the host tissues associated with high morbidity and mortality. It may affect nearly any site or organ of the body, and manifestations in the oral cavity are common. These manifestations may involve both oral mucosa and salivary glands. There are no clinical diagnostic criteria for salivary gland cGVHD (sg-cGVHD), which primarily presents as sicca syndrome with xerostomia or salivary dysfunction, similar to Sjögren’s syndrome.

Gold standard for diagnosing sg-cGVHD is through biopsy and histopathological diagnosis. There are defined diagnostic features for histopathological diagnosis of sg-cGVHD but only vague guidelines for the assessment of severity.

Standard of care can be improved by semi-quantitative grading schemes for histopathology, but also through image analysis with artificial intelligence and machine learning. Digital pathology produces large and intricate images that are not easily analysed through traditional machine learning techniques. Although modern algorithms called convolutional neural networks have improved object detection and segmentation in histopathology images, they struggle with analysis of complex context. Alternative tissue representation and machine learning adapted for analysis of relationships are emerging and may be further explored in the context of oral diseases. Using graph theory, tissue may be represented as an interconnected network of its constituents. By modelling tissue as a graph, we can also apply deep learning with graph neural networks for representation of more advanced properties of the tissue.

In Study I, we validated histopathological features in grading schemes for both diagnosis and severity rating of sg-cGVHD. We devised a semi-quantitative grading scheme based on pathology criteria defined by the National Institutes of Health’s Consensus Project and compared it to a proposed grading scheme by Imanguli et al. that was inspired by the assessment of Sjögren’s syndrome. In addition to the histopathological grading, we assessed the characteristics of the inflammatory infiltrate by quantitative immunohistochemistry through the software CellProfiler. The formalised NIH sg-cGVHD grading scheme ranged from Grade 0 to Grade IV (G0-GIV), with histopathological verification of diagnostic cutoffs for possible cGVHD at GII and Likely cGVHD at GIII. Acinar and periductal inflammation were closely correlated, and the inflammation was often widespread. The quantification of inflammatory infiltrate confirmed CD8+ T-cells to be the predominant cell type, followed by CD4+ T cells. Macrophages (CD68+) were also detected at increased levels, whereas B-cells (CD19+, CD20+) were not detected.

In Study II, we presented a model for tissue representation of oral mucosa that is based on graph theory. We proposed a 2-stage pipeline for modelling of the tissue in the form of a cellgraph with labelled edges for identification of the basement membrane. It encompassed modules for predicting cells with convolutional neural networks and constructing cell-graphs, as well as for basement membrane localisation with the deep learning algorithm GraphSAGE trained on the concept of basement membranes as crossing edges in a cell-graph. This representation yielded a model that fitted low-level features like cellular properties, their relationships in the tissue, and the high-level features of the basement membrane. The model performed with a reliable accuracy of 0.88 in predicting the basement membrane interface.

In conclusion, sg-cGVHD can be diagnosed and assessed for severity using the formalised NIH sg-cGVHD grading scheme for histopathology, and histopathological severity is associated with increased levels of CD8+ and CD4+ T cells, as well as macrophages (CD68+). Grading schemes can benefit from artificial intelligence, but this requires fair image representation. Oral tissues can be represented as graphs, and complex features like the basement membrane inferred by graph learning.

List of scientific papers

I. Tollemar, V., Arvidsson, H., Häbel, H., Tudzarovski, N., Legert, K. G., Le Blanc, K., Gunnar Warfvinge & Sugars, R. V. (2023). Grading of minor salivary gland immuno-histopathology post-allogenic hematopoietic cell transplantation. Heliyon, 9(4), e15517.
https://doi.org/10.1016/j.heliyon.2023.e15517

II. Nair, A.*, Arvidsson, H.*, Gatica V, J. E., Tudzarovski, N., Meinke, K., & Sugars, R. V. (2022). A graph neural network framework for mapping histological topology in oral mucosal tissue. BMC bioinformatics. 23(1), 506. *contributed equally
https://doi.org/10.1186/s12859-022-05063-5