Early Diagnosis And Accurate Lesion Segmentation Of Gastric Cancer Based On Deep Learning

“Development Plan of Artificial Intelligence for the Next Generation” was formulated by the Chinese government,that emphasizes the advancement of artificial intelligence in the medical health sector and the establishment of a precise intelligent medical system.This suggests that artificial intelligence will make remarkable strides in the medical field.In 2020,there were approximately 19.3 million new cancer cases and nearly 10 million cancer-related deaths in worldwide,according to statistics from the International Agency for Research on Cancer.Gastric cancer accounted for 5.6% of new cancer cases,ranking fifth,and7.7% of cancer-related deaths,ranking fourth.Relevant studies have shown that gastric cancer,as a common gastrointestinal malignancy,has a poor overall prognosis with a 5-year survival rate of only 10%-30%.The 5-year survival rate of patients with advanced gastric cancer is even lower than 10 %,while the 5-year survival rate of patients with EGC is more than 70 %-90 %.Therefore,the early detection and treatment of gastric cancer are crucial to save the lives of patients.White light endoscopy is a routine method for the detection and diagnosis of early gastric cancer(EGC),however,its accuracy(Acc)relies heavily on the professional knowledge and work experience of endoscopists.Even experts have missed or false detection.In addition,the huge workload of medical image analysis also affects the diagnostic results of endoscopists.Therefore,the ACC of EGC detection with white-light endoscopy is only 70%-80%.Based on the above problems,this study explores the detection of EGC based on deep learning methods and gastroscopic images.The aim is to use deep learning techniques improve the detection rate of EGC and the accuracy of lesion segmentation,thereby assist doctors in their work.The main innovative contributions of this paper are as follows:(1)A gastroscopic image dataset for EGC was constructed.Currently,publicly available gastroscopic image datasets primarily consist of gastrointestinal polyp images,and there is a lack of EGC-related gastroscopic image data.Therefore,the EGC images used in this study were independently collected,totaling 1,120 images.The dataset annotation process was strictly completed according to established guidelines and procedures under the guidance of professional doctors and reviewed by experts to create a EGC dataset that meets experimental requirements.Additionally,a publicly available polyp dataset,Kvasir-SEG(which includes 1,000 images),was also utilized in this paper to assist in validating the performance of the deep learning model and provide an objective evaluation.(2)A U-Net model was built and improved to enhance the accuracy of segmentation for small and/or hidden lesions.A U-Net model with VGG16 as its backbone network was constructed and trained using the EGC dataset.After testing,it was found that the model had insufficient segmentation capabilities for smaller or more concealed lesions.And then,an improved featureenhanced U-Net model(FEU-Net)was proposed.To improve the multi-scale feature extraction ability of the model and the ability to segment small targets,a hybrid convolution structure has been added,which is composed of multi-scale convolution and hole convolution.Experimental results showed that the performance of the FEU-Net model was superior to the U-Net model in segmenting EGC lesions in gastroscopic images,particularly in improving the segmentation accuracy of small and hidden lesions.(3)The Mask R-CNN(MR-CNN)model was improved to enhance EGC detection rate and lesion segmentation accuracy.Although the FEU-Net demonstrated good performance,it lacked detection capabilities and could only perform lesion segmentation.Additionally,it had limited capability in segmenting EGC lesions with complex backgrounds.To address these issues,further research was conducted based on the MR-CNN.To improve the ability of deep learning in EGC detection and lesion segmentation in gastroscopic images,an Improved Mask R-CNN(IMR-CNN)model was proposed.It is formed by adding both a Bi-directional Feature Extraction and Fusion module as well as a Purification Mechanism for Feature Channel and Space based on MR-CNN.The experimental results show that the values of Precision,Recall,Specificity and F1-Score are92.9 %,95.3 %,92.5 % and 94.1 %,respectively.These indicators for IMR-CNN model are higher than that for the original MR-CNN model,proving that our proposed IMR-CNN model has a better ability for detection of EGC and segment the lesion infiltration area.So,IMR-CNN model can effectively assist doctors in the diagnosis of EGC from gastroscopic images.