Image-based model for quantitatively characterizing spatial heterogeneity of hepatocellular carcinoma and predicting prognosis

Abstract

The application relates to a hepatocellular carcinoma tumor spatial heterogeneity quantitative characterization and prognosis prediction method and device based on an image base model, which comprises the following steps: training a hepatocellular carcinoma image base model by combining elastic deformation enhancement and a self-supervised learning framework of random masking, so as to automatically extract high-quality deep features. The model is used to extract tumor image block features, the features are reduced in dimension, and the tumor is divided into different sub-regions through two-stage clustering. A comprehensive tumor spatial heterogeneity index which integrates the aggregation, boundary mutation degree and distribution disorder degree is calculated based on the spatial distribution of the sub-regions. Finally, the index is combined with clinical characteristics, and input into a prognosis prediction model to obtain a risk score. Compared with the prior art, the hepatocellular carcinoma image base model is trained by adopting self-supervised learning, and traditional artificial defined features are replaced, so that more accurate deep features can be automatically extracted, and finally the tumor spatial heterogeneity quantification and prognosis prediction performance are improved.

Images