Research On Malignant Risk Stratification Of Thyroid Nodules Based On Multi-Label Object Detection

With the clinical application of the computer-aided diagnosis system based on deep learning,the efficiency of clinical diagnosis has been effectively improved.In the process of clinical diagnosis of thyroid diseases,the use of C-TIRADS thyroid nodule malignancy risk stratification guidelines to diagnose nodule characteristics has greatly promoted the standardization and homogeneity of the diagnostic process.But there is still a lack of research on the auxiliary diagnosis of thyroid nodules combined with the C-TIRADS standard.The auxiliary diagnosis combined with the C-TIRADS standard is more in line with the current domestic clinical diagnosis process,and more detailed diagnosis results can be obtained,which can more effectively assist doctors in diagnosis.Therefore,this thesis mainly studies how to use deep learning technology to automatically detect thyroid nodules in ultrasound images and classify the detected nodules for malignant risk,and proposes a multi-label object detection model based on the C-TIRADS standard.The main content and phased results of the research are as follows:First,a benchmark model is selected for the construction of a multi-label thyroid nodule target detection model,and the Mask R-CNN model with the best effect on nodule detection is selected as the benchmark model through comparative experiments.Secondly,the feature extraction network was improved for the benchmark model to solve the problem of poor nodule recognition accuracy.The Res Net152-FPN model was used as the feature extraction network of the improved model to improve the feature extraction ability of the model.Then,in order to enable the model to predict detailed pathological features,a multi-label detection head is designed in combination with HMCN,and the classification performance of the prediction branch is improved by enhancing constraints.Afterwards,Combined with clinical prior knowledge,the anchor box size and ratio of the improved model are customized to improve the detection accuracy of the model.Finally,the model is trained with a custom migration learning strategy to improve the model detection performance.Through training and testing the improved model on real dataset,it is verified that the improved model has improved the detection and recognition effect of thyroid nodules.The improved model achieves 94.6% recognition accuracy for thyroid nodules under the condition of an IOU threshold of 0.5,and the average recognition accuracy for various pathological features reaches 88.6%.