Spatial interrogation of tumour microenvironment using artificial intelligence

The tumour microenvironment can provide critical information for disease diagnosis, treatment planning, and prognosis. However, the complexity of its morphological, cellular, and spatial architecture hinders accurate evaluation and quantification. Deep mining of its content using knowledge-driven artificial intelligence methods can significantly benefit clinicians and patients by uncovering new disease biology and generating objective assessments in the decision-making process. In this PhD thesis, we developed deep learning based image analysis pipelines to spatially interrogate the role of the tumour microenvironment in various cancer types, including follicular lymphoma, multiple myeloma, and ductal carcinoma in situ, using multispectral immunofluorescence (MIF), multiplex immunohistochemistry (MIHC), and hematoxylin and eosin (H&E) tissue staining technologies. Firstly, we developed new deep learning-based pipelines to detect and classify single cells and segment different tissue compartments on MIF and MIHC images. The deep learning models were trained and validated using expert pathologists' annotations. Secondly, we developed tissue morphology and single-cell spatial analysis methods tailored to the tissue structures' complexity to identify spatially resolved phenotypes and spatial topography of cells to detect disease prognosis. We showed the significance of the architectural distribution of tumour infiltrating lymphocytes (TILs) on the prediction of disease outcome. Finally, we implemented an automated TILs scoring pipeline from H&E images that account for ductal carcinoma in situ spatial infiltration pattern and mimic pathologists' TILs scoring procedure. The spatial scores were associated with patient response to treatment and risk of recurrence. In conclusion, we built new deep learning based image analysis pipelines that dissect tissue structures and spatially map cell phenotypes in histopathology images and identified novel spatial prognostic features in multiple cancer types. Once validated, these methods could be utilised in clinics as decision support for the diagnosis and prognosis of cancer patients for precision medicine