Study On Ultrasonic Hyperparathyroidism Detection Method Based On DETR

Hyperparathyroidism is an abnormal condition caused by excessive activity of parathyroid glands,which damages human health,can cause tissue calcification,forms kidney stones,and can cause cardiomyopathy,hypertension and other diseases in severe cases.At present,in the routine clinical initial diagnosis,cervical B-type ultrasound imaging is the main means of screening for abnormal parathyroid function,but due to the limited clarity of ultrasound imaging,the variable characteristics of parathyroid glands and the significant individual differences of patients,only experienced professional doctors can accurately achieve the diagnosis of hyperparathyroidism,but the missed detection rate is still high.Based on the Deformable DETR(Detection Transformer)object detection algorithm,this paper innovates and improves the detection network model,and applies it to the automatic detection of hyperparathyroid lesions,in order to improve the accuracy of diagnosis of this disease through computer-aided diagnosis.The main research contents are as follows:First,the ultrasound image data of the neck parathyroid gland were statistically analyzed,and the scale,area and aspect ratio distribution of the target of the detection lesion were looked for,and multiple versions of the ultrasound hyperparathyroidism dataset were made by a variety of data enhancement methods,and the optimal version was selected according to the experimental results.Secondly,in order to improve the performance of Deformable DETR algorithm in the detection of hyperparathyroidism,the attention residual structure is introduced to improve the feature extraction network containing residual connections.In this paper,multiple feature extraction networks and multiple types of attention structures are selected and tested on the hyperparathyroidism dataset,and the test results show that the CSPDarknet-SPP feature extraction network using channel attention residual structure has achieved better performance indicators,which effectively improves the detection performance of hyperparathyroidism.Thirdly,in order to further improve the performance of the attention residual structure,this paper analyzes and redesigns the modeling method of channel information and attention weight in channel attention,and proposes a new channel attention module SE-DConv.By directly and independently modeling channel information,this module significantly reduces the attention complexity of SEnet(Squeeze-and-Excitation network)channels while maintaining high detection performance.Fourth,aiming at the problem of limited computing resources in practical application scenarios,a lightweight feature extraction network CSPDark Snet is designed by introducing operations such as shallow network space search,adaptive group convolution and adaptive group channel shuffling.The number of network parameters is only 50% of that of Res Net50,and the computational overhead is low.Using the network as a feature extraction network of Deformable DETR can obtain higher detection performance with shorter training time and lower computing resource overhead.Compared with mainstream detection algorithms,the algorithm proposed in this paper has higher detection sensitivity and can significantly reduce the missed detection rate.In this paper,a variety of improved methods for Deformable DETR object detection algorithm are proposed and applied to the automatic detection of hyperparathyroidism.Experiments have proved that the method proposed in this paper has high detection accuracy,can reduce the missed detection rate of this disease,and can assist doctors to make rapid and accurate diagnosis.