Annotations-free survival prediction with WSIs using graph convolutional neural networks.

e16501 Background: Survival prediction of cancer patients has always been an challenging problem.Tumor microenvironment (TME) Analyzation based on whole-slide-images (WSIs) has provide an effective perspective for survival prediction. However, most existing TME analyzation based on cell segmentation or classification relies heavily on labor-intensive cell-level annotations of pathologists. Furthermore, except for each individual cell or local pathological feature, survival prediction also involves local-level pathological features interactions in tumor microenvironments. This requires context-awareness based on histological features to fully infer the patient's survival risk. Therefore, we explored a model based on graph convolutional neural networks (GCNN) to perform survival prediction of cancer patients using WSIs. Methods: We utilized WSIs collected from The Cancer Genome Atlas (TCGA) to develop a graph convolutional neural network for survival prediction of cancer patients. The model leverages the advantages of graph structures to autonomously learn the histopathological contextual features in WSIs, and therefore can incorporate additional and crucial tumor microenvironment interaction information while avoiding the labor-intensive annotations. WSIs of two different cancers, KIRC and LUAD, were randomly divided into training, validation, and testing sets in a ratio of 7:1:2. The performance of the constructed model is evaluated using the test set and the results are compared with other state-of-art methods. Results: Our work is compared with other state-of-art weakly supervised learning methods for survival prediction in computational pathology. Abundant experimental results shown that our method outperformed previous methods on these two cancer types (achieving a 2.9% improvement compared to Multiple Instance Learning (MIL) and a 2.6% improvement compared to Attention MIL), with an overall c-index of 0.646. Additionally, we evaluated the interpretability of our model through attention heatmaps of low-risk and high-risk patients. Conclusions: We have developed a GCNN based model, combined with attention mechanisms, to learn features of heterogeneous tumor microenvironments and their contextual information from memory-efficient representations of highly correlated image patches for cancer patients survival prediction. This model is applicable to any weakly supervised learning task in computational pathology that involves slide-level or patient-level labels, making it an effective supplementary diagnostic tool for oncologists and pathologists.