Research On Computer-aided Diagnosis Algorithm Based On Time-Phase Contrast-enhanced Ultrasound Image

Contrast-enhanced ultrasound is a new type of ultrasound technology.It uses microbubble ultrasound contrast agents to highlight microvessels and macrovessels through strict intravascular injection.This allows quantitative or qualitative methods to evaluate vascularization in local areas.It overcomes some limitations of conventional B-mode ultrasound imaging on liver imaging,and is widely used for the detection and diagnosis of various organ tumors such as insulin,breast,and prostate.The localization of the region of interest in the contrast-enhanced ultrasound image is an important task.Doctors need to observe the morphology of the lesion in the region of interest to make the corresponding diagnosis.The biggest difference between contrast-enhanced ultrasound and conventional B-mode ultrasound is that it needs to be diagnosed by observing the change of blood vessel morphology after injection of contrast agent,which requires observing image information based on time changes.Time-phase video data also brings challenges to its application.In recent years,multidisciplinary integration has become a trend,and the cross-links between the medical field and the computer field have become closer.In particular,the field of artificial intelligence represented by deep learning has developed rapidly and has become the mainstream technology for computer-aided diagnosis.Thus the research on computer-aided diagnosis algorithm of time-phase contrast ultrasound images is mainly carried out as follows:(1)In this paper,the object localization algorithm based on weakly supervised learning is used to realize the object location of the region of interest in the ultrasound contrast image when only the image label information is used and no position information is used.The weakly supervised object localization algorithm based on deep learning is introduced,and on this basis,the AEN algorithm is proposed,which uses real ultrasound contrast data to quantitatively and qualitatively analyze its effect.The aim is to assist doctors in observing regions of interest and reduce data labeling workload of ultrasound doctors by object localization based on weakly supervised learning.(2)In this paper,we use phase contrast ultrasound video to complete the diagnosis of benign and malignant tumors.This paper introduces multiple ultrasound contrast video classification algorithms and proposes the MFSE algorithm based on them.Experiments were performed on a true contrast-enhanced ultrasound data set,and the superiority of the MFSE algorithm proposed in this paper in the diagnosis of benign and malignant tumors in temporal contrast-enhanced video was verified.The results obtained can provide doctors with reliable diagnostic recommendations.(3)At present,in two-dimensional images,using natural image pre-trained models to transfer to medical images has been verified to be effective,but few people have performed such verification in videos.This paper uses natural video pre-training models to transfer to medical video diagnostic tasks.Through experiments,it is verified that the transfer learning of natural videos to medical videos is also effective in videos.