请选择 进入手机版 | 继续访问电脑版
  • 设为首页
  • 点击收藏
  • 手机版
    手机扫一扫访问
    迪恩网络手机版
  • 关注官方公众号
    微信扫一扫关注
    公众号

apopli/diabetic-retinopathy: Diabetic Retinopathy: Segmentation, Grading and Loc ...

原作者: [db:作者] 来自: 网络 收藏 邀请

开源软件名称(OpenSource Name):

apopli/diabetic-retinopathy

开源软件地址(OpenSource Url):

https://github.com/apopli/diabetic-retinopathy

开源编程语言(OpenSource Language):

Python 100.0%

开源软件介绍(OpenSource Introduction):

Overview

This repository provides source code, submitted papers and demo for Diabetic Retinopathy: Segmentation, Grading and Localization with IDRiD dataset.Our method won the 1st place in Fovea Localization with overall 3rd place in the Localization sub-challenge of IDRiD Grand Challenge. Besides, we secured 5th rank in Segmentation of Hard Exudates.

Lesion Segmentation

We developed a data-driven method for automatic detection of retinal lesions based on their characteristics in fundus images to treat Diabetic Retinopathy. First we preprocessed retinal images to reduce image noise and did several data augmentations to make data variety more richer and distinct. After that, we segmented them using a model with UNet architecture and weighted cross entropy loss. The Unet architecture includes a shrinking path to capture context of surrounding and a symmetric expanding path that enables accurate localization. We improved and extended this architecture such that it works with very few training images and produces more accurate lesion masks. We have trained four distinct binary models for the four lesions - hard exudates, hemorrhages, microaneurysms and soft exudates.

Disease Grading

The retina images for disease classification were obtained from Kaggle dataset. The labels were provided by clinicians who rated the severity of diabetic retinopathy in each image on a scale of 0-4. The images were preprocessed, downsampled and augmented before feeding them to a 50 layer deep ResNet. Our networks achieved very competitive Kappa score of 0.74 on Kaggle Private Leaderboard, along with sensitivity of 82% and specificity of 84%.

Localization

The model consisted of two subsequent approaches to do localization. The initial model was a Convolution based model to get a tentative position of the Fovia and the Optical Center. The next model worked on a patch around the predicted value of the previous model to improve the accuracy. The motivation for sucha step was to first get a global view of the image and then focus on a local area for predicting the coordinates. The first model is a standard CNN while the second model implements UNet architecture.


The repository also provides short papers submitted to IEEE ISBI as part of the IDRiD Grand Challenge.

We also developed a Django webserver with a basic user interface to upload and test retinal images on our models. The Django site was hosted on our workstation. View demo video: https://apopli.github.io/dr_demo.mp4




鲜花

握手

雷人

路过

鸡蛋
该文章已有0人参与评论

请发表评论

全部评论

专题导读
热门推荐
阅读排行榜

扫描微信二维码

查看手机版网站

随时了解更新最新资讯

139-2527-9053

在线客服(服务时间 9:00~18:00)

在线QQ客服
地址:深圳市南山区西丽大学城创智工业园
电邮:jeky_zhao#qq.com
移动电话:139-2527-9053

Powered by 互联科技 X3.4© 2001-2213 极客世界.|Sitemap