A robust multi-modal survival prediction system for cancer patients

Abstract

This invention discloses a robust multimodal survival prediction system for cancer patients, comprising: a data preprocessing module for acquiring pathological images and genomic data and preprocessing them to obtain microscopic image package representations and genomic package representations; an image compression sensing module for extracting macroscopic image package representations from the microscopic image package representations using clustering and perceptron mechanisms; a genomic package synthesis module for reconstructing genomic package representations containing multiple biological functions using an improved denoising network, guided by the macroscopic image package representations; and a multi-granularity adaptive fusion and prediction module for achieving multi-granularity collaborative alignment and fusion of genomics and microscopic / macroscopic features based on the data missing state, and outputting prediction results. This invention uses pathological images and genomic data, and by mining dual-scale image features and reconstructing missing genomic modalities, significantly improves the prediction accuracy and robustness of the system in various clinical scenarios.

Images