Fundus image classification system based on integrated deep learning

Abstract

The embodiment of the invention provides a fundus image classification system based on integrated deep learning. The fundus image classification system comprises a pre-diagnosis classification network, a segmentation network and a final diagnosis module. The pre-diagnosis classification network obtains an initial diagnosis result based on the global information of a target fundus image; the segmentation network performs image segmentation on the optic disc, optic cup and optic nerve fiber layer states of the target fundus image based on the initial diagnosis result; and the final diagnosis module extracts a vertical cup-to-disk ratio and an ISNT score based on a result of optic disk, optic cup and optic nerve fiber layer state segmentation, and acquires and displays a final category of thetarget eye fundus image based on the vertical cup-to-disk ratio, the ISNT score and an optic nerve fiber layer defect state. According to the embodiment of the invention, firstly, glaucoma pre-diagnosis is carried out on a target fundus image, and a proper target segmentation network is selected, so that the segmentation precision is improved; and when glaucoma judgment is carried out, the accuracy of a classification result is further improved by combining a plurality of quantitative indexes capable of reflecting the disc edge form.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a fundus photography classification system based on integrated deep learning. Background technique [0002] In the field of AI diagnosis of glaucoma based on fundus photography, there are mainly two types of deep learning methods: deep classification network and deep segmentation network. Deep classification networks are a class of end-to-end glaucoma diagnostic networks that can extract features from fundus photographs and make glaucoma diagnoses. However, the features extracted by such networks and their diagnostic logic are opaque to doctors. In addition, the training of such networks requires tens of thousands of training samples. [0003] Deep segmentation networks are a class of image segmentation networks that can segment the optic disc and optic cup from fundus photographs. Furthermore, a vertical cup-to-disk ratio (VCDR) is calculated and a glaucoma diagnosis is mad...

AI Technical Summary

Problems solved by technology

This type of method has three limitations: firstly, the optic cup segmented by this kind of network still has a large deviation from the optic cup marked by the doctor; secondly, only using VCDR is not enough to reflect the complete information of the shape of the disc edge; finally, there is no use other than the optic disc Information

As a result, the accuracy of using AI to diagnose glaucoma in the prior art is not high

Images