A Study On Real-time Semantic Segmentation Of Thyroid Nodules Based On Fully Convolutional Network

Thyroid nodule is a very common type of thyroid disease in clinical practice,and most of them are benign.However,its potential malignancy always endangers the health of patients,so the early diagnosis and treatment of thyroid nodule are very important.Ultrasonography is the preferred imaging method for the diagnosis of thyroid nodules.In recent years,some machine learning methods have been used in computer-aided diagnosis(CAD).Computer-aided diagnosis requires high accuracy and real-time performance to effectively improve the diagnostic efficiency of doctors.The existing semantic segmentation algorithms based on deep learning have high accuracy but insufficient real-time performance.In this paper,a real-time architecture of medical image semantic segmentation is proposed,which improves the segmentation efficiency while maintaining high accuracy.In particular,we use dense connection,dilated convolution and factorized filters to design dense dilated block,so that the network remains efficient while retaining remarkable accuracy.We use dense connectivity to reduce the vanishing gradient problem,optimize feature reuse,and redesign layers using dilated convolution and factorization convolution to maintain significant accuracy.In addition,aiming at the class imbalance problem in medical image binary segmentation,a loss layer optimization method is proposed from the perspective of pixels,which further improves the accuracy of the network.Testing a single image,the proposed method has an inference time of approximately 7.92 ms on a single TITAN Xp GPU,and approximately 0.63 s on the Core(TM)i5-4590 CPU @ 3.30 GHz CPU.Therefore,the proposed method can also meet the requirements of real-time segmentation on common configuration equipment.A series of experiments based on Thyroid dataset show that our approach has achieved an accuracy nearly as high as that of the state of the art like CE-Net,while our inference speed is more than 6X faster than top-accuracy approaches.The proposed architecture runs fast and occupies less memory.It is an ideal choice for mobile and embedded devices.