Automatic Thyroid Nodule Detection On Ultrasound Images

Thyroid carcinoma is one of the most common endocrine diseases in adolescents,which is a serious threat to people’s health.Statistics show that the incidence of thyroid nodules has significantly increased over the past decades.Accurate location and determination of thyroid nodules is the key point in early diagnostic practices.Ultrasonography is the most widely employed imaging method due to its avoidance of ionizing radiation,real-time visualization,relatively inexpensive price,non-invasive diagnosis and high sensitivity in the process of thyroid scanning.However,it is a laborious and time-consuming task for radiologists to keep scanning thyroid glands and reading thyroid ultrasound images for lesion location during clinical examinations,which may lead to subjective interpretations and inter-observer variabilities.As a result,manually assessment of thyroid nodules is often not robust,thus the computer-aided detection(CAD)system has become one of the major research subjects in medical imaging over the past years.Develop a precise ultrasound nodule detection method with computer to avoid the influence of subjective factors and reduce the burden on doctors is a main topic in computer-aided thyroid diagnosis.However,lesion regions in ultrasound images are usually blurred,vague in margin and irregular in shape.In addition,ultrasound images are commonly monochrome and have various inherent artifacts and noises,such as speckle noise and intensity inhomogeneity.To overcome the difficult points mentioned above,in our study,we propose a faster and more accurate thyroid nodule detection network based on the Faster R-CNN(Region-based Convolutional Network)framework by adding some strategies including feature pyramid,spatial remapping and anchor-box redesign.Quantificational results indicate that our method has shown advantages over the original Faster R-CNN,which can locate the target nodules accurately with a m AP of 92.79% and the detection rate of 15 fps.By comparing the performance to senior doctors,we find that the proposed method has the ability to detect thyroid nodules in ultrasound images more rapidly and accurately than manual work,which indicate the value of our method in assisting thyroid image interpretation.In this paper,we have proposed an improved nodule detection framework based on Faster R-CNN,detecting thyroid nodules in ultrasound images accurately and efficiently.This automatic detection method has the potential to assist inexperienced community doctors in accurate and rapid thyroid nodule diagnosis and to place less burden on senior physicians.According to the clinicians,as long as the corresponding image set is available for training,this method is theoretically effective for the ultrasound image understanding of other superficial organs which have similarity with thyroid concerning the tissue characterization and structure.