A Deep Convolutional Neural Network Approach for Predicting Cumulative Incidence Based on Pseudo-Observations

• Activity Number: 132 - SLDS CSpeed 1
• Type: Contributed
• Date/Time: Monday, August 9, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Pablo Gonzalez Gonzalez Ginestet* and Philippe Weitz and Erin Gabriel and Mattias Rantalainen
• Companies: Karolinska Institutet and Karolinska Institutet and Karolinska Institutet and Karolinska Institutet
• Keywords: Convolutional Neural Network; Machine Learning; Survival Analysis; Pseudo-Observations

Abstract

The use of medical images in precision medicine has recently become an important tool to improve the prognosis of a patient or predict the future course of a patient disease. Convolutional neural network (CNN) is at the forefront of image analysis and has become the state-of-the-art on image-based precision medicine. CNN has been used to address the task of predicting time-to-event outcomes from whole-slide images. These prior works combined modern CNN architectures with Cox regression for prediction of time-to-event outcomes, keeping the assumption of proportional hazard. In this paper, we propose an approach based on pseudo-observations for predicting time-to-event outcomes from medical images using any CNN architecture. We expect in that way to avoid the proportional hazard assumption and any special tailored loss function to handle right-censoring. We assess the performance of the proposed methodology through simulations and we compare it to existing methods. We illustrate the proposed method using whole-slide images of breast biopsies and clinical structured data obtained from The Cancer Genome Atlas Breast Invasive Carcinoma(TCGA-BRCA).