Prediction Of Thyroid Tumor Disease Based On Pathological Image Analysis

As society continues to develop,people’s lifestyles are also changing.Influenced by factors such as diet,sleep,and environmental pollution,the incidence of thyroid cancer in China has been increasing rapidly in the past decade,making it the fastest-growing solid tumor in terms of incidence rate,and it also has a relatively high probability of metastasis.The diagnosis of thyroid tumors is based on thyroid ultrasound images,but the accuracy of ultrasound diagnosis of whether the tumor has lymph node metastasis is relatively low.Therefore,this thesis aims to develop a system that can segment thyroid lesions and predict lymph node metastasis,which can be used to evaluate the patient’s condition and provide a reference for the doctor’s diagnosis,thus improving the accuracy of the diagnosis.The main achievements of the thesis are:1.Dataset construction:This thesis constructed an ultrasound dataset of papillary thyroid carcinoma,and the labeling content of the dataset was annotated by professional ultrasound physicians.According to the doctor’s requirements,a labeling tool was developed and the initial data provided by the doctor was processed twice to form a dataset that meets the requirements of this thesis,including a total of 1980 ultrasound images.2.this thesis designed a complete dataset preprocessing function:Using the MMDetection detection network,the effective area of the annotated ultrasound images by doctors was extracted,and the data augmentation method was used to solve problems such as low contrast of ultrasound images and difficulty in distinguishing thyroid nodules,thus improving the usability of the dataset and providing good dataset conditions for the image classification and segmentation tasks of this thesis.3.R-DUNet multitasking model:ResNet is a commonly used backbone network with significant effects in feature extraction.U-Net can be trained from small datasets,solving the problem of not having a large dataset.This thesis proposes a multi-task learning-based image segmentation classification algorithm,which is based on the design ideas of ResNet and U-Net network structures,and constructs the R-DUNet model.The experimental results show that the R-DUNet model constructed in this thesis can efficiently segment thyroid nodules and predict whether papillary thyroid carcinoma has lymph node metastasis,which can assist doctors in clinical diagnosis and provide convenience for doctors in the diagnosis of thyroid tumors.