Ultrasound Images Classification And Object Detection Based On Deep Learning

Ultrasound is widely used in the clinical diagnosis of ocular diseases.B-type ultrasound is mainly used in the field of ophthalmology.Ocular ultrasound images can show the disease and its shapes in the eye,including intraocular and periocular areas such as the vitreous,lens,retina,etc.In this paper,seven kinds of ultrasound images are classified,namely normal images and vitreous opacity,hemorrhage origanization,choroid detachment,retinal detachment,stellate degeneration,vitreous silicone oil filling.As there are many kinds of eye diseases,which are concentrated in a small area,ultrasound images contain noise and other factors,which directly affect the doctor’s interpretation.Moreover,because the characteristics of some diseases are very similar,such as choroid detachment,retinal detachment and so on,the interpretation and diagnosis of ultrasound images will be highly dependent on the clinical experience of doctors.Usually,the doctors who perform the interpretation of ultrasonic images have many years of diagnostic experience.Therefore,this paper uses the deep learning method to automatically classify ocular ultrasound images,and realizes the purpose of diagnosing which diseases patients belong to by directly classifying ultrasound images.The conversion from the complete diagnosis of doctors to the diagnosis results for doctors’ reference is of great significance for improving the diagnosis efficiency of doctors and reducing the pressure of doctors.This paper uses deep learning method to classify 7 kinds of ultrasonic images.For this purpose,this paper introduces a data set containing box annotations and category annotations.The data set is historical eye diagnosis data from 2016 to 2021,collected directly from eye hospitals and annotated by professional ophthalmologists.In order to improve classification accuracy,this paper added YOLOv5 s network for object detection before direct classification,and designed a two-stage experimental framework for classification.First,YOLOv5 s was used to detect the target region and eliminate the interference of non-target region in the image.Then Res Net101 was used to classify the region detected by YOLOv5 s.To solve the problem of insufficient data samples and uneven distribution of image quantity,this paper expands the data by data enhancement method to improve it.In view of the difficulty in distinguishing the features and their similarity,this paper replaced the fine-grained loss function and constrained the Res Net101 network from ordinary image classification to fine-grained classification,so that the network could pay more attention to the key features that could represent each category.YOLOv5 s finally reached 89.0%m AP,and the classification accuracy of Res Net101 reached 81.0%.After data enhancement and training with Focal loss and MCloss,the classification accuracy increased by 4.4% to 85.4%.