OCT Image Based Generation Of Retinl Adiagnostic Report And Construction Of Pathological Information Dataset

The rapid rise of the internet and all kinds of electrical products have changed the way people live and work,and also led to very serious eye diseases,such as pathological high myopia,glaucoma and other irreversible eye diseases.Optical coherence tomography(OCT)is one of the most commonly used imaging techniques in ophthalmologic clinical that requires experienced clinicians to review and diagnose the images.With the increasing of patients with eye diseases,the workload of OCT reading is increasing rapidly,which becomes a huge burden for the ophthalmologists.It is urgently to develop the automatic generation of retinal diagnostic report based on OCT images to assistant doctors in reading and diagnosing.Due to the noisy OCT images,the blurred boundaries of lesions and the small proportion of key information,the automatic generation of retinal diagnostic report based on OCT images is extremely challenging.In this thesis,the automatic generation technology of retinal diagnostic report based on OCT images is mainly studied by combining the methods of long short-term memory(LSTM)network,multi-scale feature fusion and attention mechanism.The construction of standardized pathological information description dataset is also studied.The main work and innovations of this thesis are summarized as follows.An automatic generation algorithm for OCT image diagnostic report based on multiview and multi-scale feature fusion is proposed.Combining with LSTM network and attention mechanism,the feature maps are extracted from two retinal OCT images of different views via weight sharing strategy respectively and fused at different stages of the network.At the same time,based on multi-scale feature fusion method,a weight map is generated using deep features,which can guide shallow network to focus on objects of interest.The proposed generation method is evaluated on our retinal OCT image dataset and reaches a bilingual evaluation understudy(BLEU)of 0.44.In addition,ablation and comparison experiments are designed to verify the effectiveness of the proposed multi-scale feature fusion network.A complex sentence segmentation method based on dependency parsing is proposed,which uses clinical retinal OCT image reports to construct the standardized pathological information description dataset.Due to the numerous clinical reports and the meaningless contents,the construction of the standardized pathological information description dataset is a huge task.The text processing steps proposed in this paper are as follows:1)Jieba module is used for text cleaning and abnormal words screening.The text preprocessing is performed according to the abnormal words.2)Two types of complex sentences(multi-subject sentences and multi-predicate center sentences)are divided into simple sentences with single syntactic structure.First,sentence dependency parsing is performed to extract the dependency relations between words.Then,the effective components of sentences are extracted,classified and combined based on the dependency relations.3)The OCT images are screened for the construction of a high-quality retinal diagnostic report dataset.An automatic diagnostic report generation system for retinal OCT images is designed and implemented.The system interface adopts the client-server architecture.The algorithm is deployed on a server with graphics processing unit(GPU).The client interface communicates with the server through hyper-text transfer protocol(HTTP),invokes the server algorithm,and finally displays the results on the interface.