SELF-SUPERVISED LEARNING GUIDED TRANSFORMER FOR SURVIVAL PREDICTION OF LUNG CANCER USING PATHOLOGICAL IMAGES

Accurate overall survival (OS) prediction for lung cancer patients is of great significance, and the histopathology slides are considered the gold standard for cancer diagnosis and prognosis. However, the current methods usually lack extracting effective features and ignore the utilization of spatial information. To address these challenges, we propose a self-supervised learning guided transformer framework (SET) for OS prediction with whole slide images (WSIs). We introduce self-supervised learning to exploit the characteristics of pathological images and thus capture domain-specific contextual representations. Furthermore, we design a dualstream position embedding architecture to facilitate aggregating global spatial information. The experimental results on the lung cancer dataset of stage III-N2 demonstrate that our proposed algorithm can achieve a better concordance index compared with state-of-the-art methods. Moreover, the proposed method can significantly divide patients into high-risk group and low-risk group to assist the personalized treatment.