Abstract

Info

Author(s):: LUO Siyan1; WU Doudou2; LIU Qianwei2; WANG Xinzhou3; RAO Xiangrong1; 1. Department of Nephrology, Guanganmen Hospital, China Academy of Chinese Medical Sciences, Beijing 100053, China 2. Department of Dermatology, China-Japan Friendship Hospital, Beijing University of Chinese Medicine, Beijing 100029, China 3. College of Electronic and Information Engineering, Tongji University, Shanghai 201804, China

Keywords:: Keywords: machine vision deep learning tongue diagnosis spot detection

Abstract:: Abstract: Objective To identify the distribution of spots in different areas of tongue intelligently. Methods After the initial detection of tongue spots by convolving the tongue image with LoG operator, the labeling of spots were fine-tuned by human interaction, and the convolutional neural network model (Fast-RCNN) was trained with the labeled dataset. Results With the 240 images collected by the instrument for tongue image as training set and 60 images as test set, a recall rate of 90.78% was obtained, indicating that the proposed method was superior to the existing methods. Conclusion The data pre-labeling and manual fine-tuning method proposed in the study makes it possible to label fine-grained spots. Based on the dataset accurate to a spot, convolutional neural network is introduced to detect the spot distribution at sub-pixel level, and the results can provide objective, quantitative and automatic reference for clinical diagnosis of traditional Chinese medicine.