Study of tumor tongue image recognition via convolutional neural network and transfer learning
Objective To explore the performance of the convolutional neural network-ResNet152 model in identifying tumor tongue image.Methods A dataset consist of 5 943 tongue images,including 1 433 tumor tongue image and 4 510 non-tumor tongue image were collected from Dongzhimen Hos pital,Beijing University of Chinese Medicine,Dongfang Hospital Beijing University of Chinese Medicine,and Beijing University of Chinese Medicine Third Affiliated Hospital from January 2019 to December 2021.After image preprocessing,1 000 tongue image were randomly selected as the test set and the remaining 4 943 as the training set.The tongue images of 4 943 training sets were amplified to 54 535 by image amplification technique,and then input into ResNet152 model of convolutional neural network pre-trained on ImageNet-2012 data set to establish an automatic tongue image recognition system.Then 1 000 tongues of test sets were fed into the model and the recognition results were recorded.Finally,the GRAD-CAM technology was used to visually analyze the tongue image correctly identified as tumor by the test set model,and the tongue image features focused on tumor tongue image were identified statistically and analytically by the model.Results In identifying tongue images of tumor,the accuracy of conv-olutional neural network-ResNet152 model was 85.7％,recall rate was 84.9％,precision was 85.5％,F1 score was 85.2％,and area under curve was 91.3％.A visual analysis was conducted on tongue image correctly identified as tumor,revealing that the tongue features with the highest contribution to the model's correct tumor recognition were ecchymosis and fissures.Conclusion The convolutional neural network-ResNet152 model provides a non-invasive and efficient approach that contributes to tumor detection and diagnosis.The features of ecchymosis and fissures might be the primary focus of the model in predicting tumors based on tongue images.
NETL
NSTL
万方数据
