Comparison of deep learning and model-based approaches for spatial profiling of the immune tumor environment on multiplex image data

Sammanfattning: The demographics of the tumor microenvironment (TME) impact the Immunotherapy responses for lung cancer patients. Given the heterogeneity of immune cells present within TME, the distribution patterns of different subpopulations of T-cells can be exploited to predict short-term or long-term survival of lung cancer patients. The aim of the present thesis is to implement a logistic regression approach for predicting patients’ survival and compare the aforementioned approach to the standard density-based statistical analysis. Moreover, an additionally developed model was applied, the Graph convolutional neural network method, a deep learning approach that considers the spatial distribution of the investigated immune components. Following model implementation and efficacy comparison, we found that the convolutional neural network model exhibited the highest accuracy of 70% in differentiating between the two survival groups. At the same time, the statistical analysis relied on many assumptions of the lymphocyte cells with the highest density that can define the characteristics of the long-term survival group.