Abstract
Soft tissue sarcoma (STS), which is a rare cancer, arise in soft tissues in the body, and it is difficult to diagnose and treat in general hospitals. This study focuses on obtaining a quantitative index that supports doctors building a treatment plan for STS based on the patient's risk of experiencing clinical events in the future. This index is expected to reduce the burden on doctors to make decision and realizing the tailored treatment considering the patient's condition. This paper proposes a survival period prediction method using convolutional neural network (CNN) and pathological images. The number of subjects is not enough to train a complex CNN model for regression because the STS is one of the rare cancers. We employ the mean-variance loss to perform the regression using a classification CNN called ResNet18. Furthermore, we propose the fit-label to efficiently train the classification CNN for the regression task. Experiments on 44 whole slide images of 44 patients showed that the proposed fit-label outperformed hard- and soft-labels, achieving the lowest mean absolute error of 6.5 months and the highest concordance index of 0.893. These results suggested that the proposed method can be used to evaluate the risk level of STS patients.
