Rearch On White Light Colonoscopy Adenomas Detection Technology Based On Deep Learning

Colorectal cancer is a very common malignant disease,The incidence rate of cancer ranks third in the world,and the mortality rate is fourth in the world.Colorectal cancer is usually caused by noncancerous neoplastic "polyps" in the colon or rectum.Some polyps will proliferate abnormally and become early colorectal cancer: adenoma.Colorectal endoscopy is a commonly used method in clinical examination for colorectal cancer,and the detection rate of adenomas is one of the criteria to measure the quality of colonoscopy.Studies have shown that it is important to improve the detection rate of adenomas under colonoscopy to reduce the mortality and intermittent incidence of colorectal cancer.At present,doctors use digestive endoscopy to determine whether it is an adenoma.The detection rate of adenomas is inevitably affected by factors such as physician experience,fatigue,and detection rate,which can easily cause misdiagnosis and missed diagnosis.Therefore,computer-aided diagnosis has become a key method to improve the detection rate of adenomas.The current research on the application of computer-assisted diagnostic techniques in colonoscopy focuses on the diagnosis of adenomas based on Narrow Band Imaging(NBI).However,the main time during the operation is still under White Light Endoscopy(WLE).It is necessary to rely on the doctor to determine the suspected case before switching the NBI image.Therefore,the NBI image diagnosis still cannot get rid of the doctor's prejudgment.However,the diagnosis of white adenomas is still missing due to the insignificant adenoma characteristics.This thesis proposes a new K-Refine Det network to complete the diagnosis of adenomas under white light,and greatly improves the detection rate of adenomas.The main research contents are as follows:1.A novel K-Refine Det model for computer assisted diagnosis of white light colonoscopic polyps and adenomas was proposed based on the characteristics of polyps/adenomas endoscopic images.This model mainly consists of three modules: the anchor refinement module(ARM)removes some negative samples in the prior boxes to reduce the search space,and roughly adjust to the prior box's position.The object detection module(ODM)uses five different size feature layers to make final judgments on the position and category of the object to increase the detection of small targets.In order to improve the detection accuracy,a transfer connection block(TCB)uses the transposed convolution method to match the size of the high-level features extracted by the ARM module and the low-level features,and achieve feature fusion by element-wise way.Finally,the converged transferred feature layer is transmitted to the ODM module.2.A dataset of colonoscopic polyps and adenomas under white light colonoscopy was constructed to learn model's parameters.In this thesis,the training dataset is mirrored,scaled and translated to increase the number and diversity of training samples,so as to improve the robustness of the model.3.The proposed K-Refine Det model was tested in the constructed dataset,and the experimental results showed that the precision,accuracy,specificity and sensitivity of adenoma images under WLE's colonoscopy reached 92.2%,93.6%,91.2% and 95.8%.The precision,accuracy,specificity and sensitivity of polyps were 95.0%,91.4%,95.8% and 86.6%.In addition,the area under the receiver operating characteristic(ROC)curve(Area Under Curve,AUC)of adenomas and polyps were reached 95.12% and 94.53%,indicating that the model could better classify adenomas and polyps and improving the detection rate of adenomas.The results presented in this article can provide a reference for colonoscopy doctors and have important significance in assisting the diagnosis of patients' conditions.