Multi-modal advanced deep learning architectures for breast cancer survival prediction

Breast cancer is the most frequently occurring cancer and has compelling contributions to increasing mortality rates among women. The manual prognosis and diagnosis of this disease take long hours, even for a medical professional. A model with better predictive power can benefit cancer patients from going through the toxic side effects and extra medical expenses related to unnecessary treatment. Medical professionals can be benefited from early-stage detection and selection of the appropriate cancer treatment plan. The availability of multi-modal cancer data, i.e., genomic details, histopathology images, and clinical details, supports the researchers in proceeding with the development of multi -modal based advanced deep-learning models. This research proposes gated attentive deep learning models stacked with random forest classifiers, which use multi-modal data and produce informative features to enhance the breast cancer prognosis prediction. It is designed as a bi-phase model; phase one uses a sigmoid gated attention convolutional neural network to generate the stacked features, while phase two passes the stacked features to the random forest classifier. The comparative study of the proposed and other existing methods over METABRIC (1980 patients) and TCGA-BRCA (1080 patients) datasets illustrate significant enhancements, 5.1% in sensitivity values, in the survival estimation of breast cancer patients. (c) 2021 Elsevier B.V. All rights reserved.