WCSE

Classification of Gastric Cancer based on Teacher-attention Distillation and Improved Dual-path Network

Abstract— Early discovery of gastric cancer plays an important role in the clinical prognosis of gastric cancer. In recent years, using deep learning to classify CT medical images effectively improves the accuracy and efficiency of classification. But traditional deep learning models lack the ability to learn rich contextual information from CT data. This paper aims to explore the way that use weakly label area image to improve the accuracy of classification of gastric CT (Computer Tomography, CT) image differentiation. We propose a new model structure that combines improved dual-path network (DPN) to reuse and mine new image features， and uses teacher attention distillation to encode rich contextual information. Combining our contributions, we are able to achieve most 0.8135 AUC. The overall results show better AUC performance than the state-of-the-art.

Index Terms— Differentiation status of gastric cancer, CT image classification, DPN, TAD

Chenchao Huang
School of Data Science Engineering, East China Normal University, CHINA
Wei Peng
Information Technology Services, East China Normal University, Shanghai, CHINA
Kun Yu
School of Data Science Engineering, East China Normal University, CHINA
Wenzhi Bao
School of Data Science Engineering, East China Normal University, CHINA

[Download]

Cite: Chenchao Huang, Wei Peng, Kun Yu, Wenzhi Bao , "Classification of Gastric Cancer based on Teacher-attention Distillation and Improved Dual-path Network ," 2021 The 11th International Workshop on Computer Science and Engineering (WCSE 2021), pp. 85-92, Shanghai, China, June 19-21, 2021.