Research On Deep Learning Method For Gastric Cancer Prediction Based On Tongue Features Of Traditional Chinese Medicine

Gastric cancer is a malignant disease that seriously endangers the health and normal life in human beings.At present,gastroscopy is the main modality used for detecting and diagnosing gastric cancer.However,due to its invasiveness,high cost,and complicated operation processes,the detection of gastric cancer through gastroscopy is limited.Therefore,proposing a new screening modality is critical in improving the diagnosis rate of gastric cancer.According to traditional Chinese medicine theory,"the tongue is the external manifestation of the spleen and stomach,and the coating on the tongue reflects the condition of gastric energy".Changes in tongue features can indicate the health condition of the stomach.The accuracy of this diagnosis is influenced by the physician’s expertise and experience.With the wide application of artificial intelligence technology in the medical field,tongue features can be detected and extracted using image processing and deep learning.The correlation between tongue features and gastric cancer is being investigated.Based on this association,a gastric cancer prediction model is built.It proves that tongue features can serve as a reliable basis for detecting gastric cancer,providing strong support for early screening of the disease.The main work of this paper includes the design of tongue image preprocessing method,the investigation of the relationship between gastric cancer and tongue features,and the design of a gastric cancer prediction model based on deep learning.(1)Design of a tongue image pre-processing methodA method for localizing and pre-extracting the tongue region is proposed to effectively reduce interference from non-tongue regions in subsequent tongue segmentation and feature analysis.Based on the Dlib library for face feature points detection,the coordinates of four key points in the clinical images are acquired,including the sides of the mouth,the philtrum and the chin.The tongue region can be localized and pre-extracted using these coordinates.A dataset for tongue image segmentation is constructed by contour annotation of the pre-extracted tongue images.And it is expanded to 4,375 sheets using data augmentation.The dataset is divided into a training set,validation set,and test set with a ratio of 7:2:1.A tongue image segmentation model based on Deeplab V3+ network is built and trained on a self-built tongue dataset.Its segmentation effect is assessed using evaluation metrics.The results demonstrate that the Deeplab V3+ tongue image segmentation model achieves an average pixel accuracy of 98.93% and an average intersection ratio of 97.96%.It has high accuracy and excellent segmentation effect.(2)Investigation of the relationship between gastric cancer and tongue featuresTongue features that have a strong association with gastric cancer are selected.The interpretability and robustness of the gastric cancer prediction model are improved.By combining physician’s guidelines with actual clinical acquisition,nine features are extracted to provide an objective analysis of tongue images in both gastric cancer patients and non-gastric cancer subjects.The associations between tongue features and gastric cancer are analyzed using statistical methods and machine learning techniques.Tongue features that are highly associated with gastric cancer are obtained by significant difference analysis,importance analysis and correlation analysis.Comparing the analysis results of the three methods,tongue shape,saliva,tongue thinness and tongue coating color are selected as the tongue features for predicting gastric cancer.(3)Design of a gastric cancer prediction model based on Efficient NetBy combining prior knowledge and deep features,a gastric cancer prediction model is constructed.A dataset is built using 703 pre-processed tongue images for gastric cancer prediction.The dataset is divided into a training set,validation set,and test set with a ratio of8:1:1.The Efficient Net-B0 is used as the backbone network to build the prediction model.The model is trained on the gastric cancer tongue dataset.And migration learning is used to enhance the model training effect.After the training,the accuracy and F1-score are used to evaluate the performance of the gastric cancer prediction model.The results show that the prediction model could accurately distinguish between gastric cancer patients and non-gastric cancer subjects with an accuracy of 93.6% and an F1-score of 92.6%.