CNN-based Prognostic Prediction of malignant Soft Tissue Tumors Using Pathological Images

Abstract

Malignant soft-tissue tumors are rare cancers with only about 3,000 patients per year in Japan. Pathological diagnosis is essential for treatment, but there are few pathologists familiar with malignant soft-tissue tumors, medical institutions specializing in their diagnosis, and diagnosis is time-consuming. We have been studying the estimation of tumor malignancy and patient prognosis using convolutional neural networks (CNNs) from pathological images. In this paper, we propose the survivavility estimation method and the survival-time prediction method for patients with malignant soft-tissue tumors using Inception v3 and ResNet14 whose convolutional layers are trained using a autoencoder. In the 4-fold cross-validationexperiment using 47 pathology images from 26 subjects, we successfully predicted the survival rate per image, non-survival rate per image, survival rate per subject, and non-survival rate per subject with F-measure of 0.847, 0.743, 0.909, and 0.824, respectively. Survival time estimation only by using non-surviving patients achieved 4.57 months error.