Font Size: a A A

Attention Model Ensemble Method For Cataract Diagnosis

Posted on:2022-08-01Degree:MasterType:Thesis
Country:ChinaCandidate:X F ZhangFull Text:PDF
GTID:2504306551470724Subject:Master of Engineering
Abstract/Summary:PDF Full Text Request
Cataract is one of the eye diseases with the highest incidence and blindness rate nowadays.For the prevention of blindness,timely detection and accurate diagnosis of cataract in the early stage is particularly important.Clinical cataract diagnosis often requires professional knowledge and rich experience.Cataract patients in poor and remote areas often miss out on the opportunity for diagnosis and treatment due to the lack of specialized ophthalmologists.Since the development of deep learning technology,it has been widely and successfully applied in many image classification tasks.However,there are still huge challenges in the automatic detection task of cataract.This is due to two characteristics of cataract and eye B-ultrasound images.First of all,the lesion area of cataract is located in the lens of the eyeball,but the lens occupies only a very small part in the B-ultrasound image of the eye.Secondly,different cataract patients have different and diverse features and areas of lens lesions.Moreover,the appearance of the cataract lens and the normal lens in the eye B-ultrasound image is very similar.This leads to very small differences between the B-ultrasound images of the cataract eye and the normal eye,showing great similarities.According to the above characteristics,this paper proposes a multi-model ensemble method based on residual attention for diagnosis of cataract.The main contributions are as follows:1.Construct an ensemble model for cataract detection,which is composed of three classification networks with the best effects and one integrated module.It is equivalent to synthesizing the judgments of multiple ophthalmologists to obtain more reliable diagnosis results.2.Use the target detection network to locate the eyeball region and remove the interference of irrelevant background in the eye B-ultrasound image.3.Design the residual attention module,then embed in the classification network to make the model pay more attention to the lens and the location of the disease,thereby increasing the attention weight of the lens.4.The experimental results based on the clinical eye B-ultrasound image dataset show that the proposed multi attention model ensemble method can adaptively focus on the abnormal areas of the eyeball where cataracts disease occur,and achieve 97.5% accuracy on the test set,exceeding other single model.
Keywords/Search Tags:cataract classification, B-scan eye ultrasound images, deep learning, ensemble learning, attention
PDF Full Text Request
Related items