Combination of Deep and Handcrafted Features Enabling Automated and Accurate Detections of Early Gastric Cancer Lesions in Histopathology Images
Yihao Luo · Investigación Clínica · 2019
Automatic and accurate lesion recognition of histopathology slides plays a key role in cancer diagnosis, treatment, and prognosis. Usually, CNN-based methods demonstrate strong performance in computational histopathology analysis. But a limitation of the deep-only model in the medical task was a lack of specific clinical context—heterogeneous clinical details are diminished because of the oversized kernel and premature pooling. In this work, we present a scheme combining several handcrafted descriptors for the backbone network with different improvements by the classification task of early gastric cancer in histopathology images. By adding various forms of handcrafted features to the neural network, we enhance the capacity of CNN to represent local contour and rotation characteristics of pathological tissues. The proposed framework, being the first approach to detect early gastric cancer (EGC) lesions on whole-slide histopathology images, achieves an overall 94.50% specificity and 96.35% sensitivity, where it shows improved performance and convergence than baseline models. Furthermore, we demonstrate the applicability of our method for feature representation on a subset of Imagenet which emphasizes edge and rotation characteristics.