Deep Learning Based Segmentation of Multiple Tissue Structures in Histology Images for Predictive Cancer Analytics

Abstract

According to a survey of global health metrics, cancer is one of top five leading lethal diseases around the world. It has the capability to proliferate into other parts of body and often recurs again after the treatment or removal from body. This immensely growing issue is now-a-days a hot topic among pathologists and researchers community. Among the analytic factors to study tumor aggressiveness and disease recurrence, density of micro-vessels (MVD), Lymphovascular invasion (LVI) and Perineural Invasion (PNI) are considered key prognostic factors. The manual identification of micro-vessels and nerves is time consuming, laborious and highly prone to human error. Computational pathology is an emerging field striving to improve patient care by incorporating modern algorithms to the traditional analysis procedures of microscopic slides. To overcome the challenges of multi scale, multi-shape and slight intensity variant histopathology structures, a deep neural network based hybrid semantic segmentation architecture is proposed. It comprises the fundamentals of encoder-decoder structure with the essence of parallel path network. The framework is specifically designed to improve the accuracy by focusing mega to minor object details during every block of segmentation network. The encoder uses Multi-scale feature extraction block made up of ResNeXt Blocks. This organization is effective to encode coarse to fine grained features from all specifications and dimensions while limiting the number of learnable parameters. The decoder is a combination of feature fusion and feature erudition while step by step mapping them back to pixel map. Monte-Carlo Dropout based uncertainty maps are also generated at prediction time. The proposed architecture is trained and tested on generated Nerve and micro-Vessel Semantic Segmentation Dataset (NVSSD). The trained architecture outperformed the existing state-of-the-art networks like FCN, Unet, SegNet, Deeplabv3+