Deep learning integration of diverse data types for survival prediction in serous ovarian cancer

Ovarian cancer, known for its high mortality and treatment resistance, demands precise survival predictions for personalized treatments. This research develops the Ovarian Cancer Vision Transformer Aggregated Network (OViTANet), a novel deep learning model that forecasts survival in serous ovarian cancer by combining histopathology images with clinical and genetic data, such as mutations, copy-number variations, mRNA-sequencing, DNA methylation, and protein expressions. Using data from the Cancer Genome Archive, OViTANet was developed and rigorously tested. Moreover an external validation set procured from Başkent University Hospital was used to evaluate the effectiveness of the image feature extractors in the model. The performance of OViTANet was meticulously categorized based on the data categories, each reflecting varying sample sizes due to the disparate availability of genetic data. When compared to an existing multi-modal network in the literature, OViTANet outperforms across all data categories for predicting overall survival, with concordance indices ranging from 0.5991 ± 0.0369 to 0.6902 ± 0.0499. Although the model's efficacy slightly diminished for disease-free survival prediction–with concordance indices spanning 0.5385 ± 0.052 to 0.5989 ± 0.0284–it still performed better than the comparative model across most data categories. This work not only enhances survival prediction through data integration but also contributes to understanding serous ovarian cancer's genetic traits, marking a significant advancement in oncological analysis and leading to more effective, personalized treatments to improve patient outcomes.