【深度学习】眼底图像之视盘和黄斑分割的探索

落日映苍穹つ 2022-10-21 01:26 447阅读 0赞

最强知识图谱！   学习Python或者全栈开发强烈推荐！我说的！

![在这里插入图片描述][20210525100706727.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
**针对Java开发者的图谱，一样很强。**

![在这里插入图片描述][20210525101156408.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]

## 【深度学习】眼底图像之视盘和黄斑分割的探索 ##

文章目录
    1 Optic Disc 数据集
    	1.1 ORIGA-650
    	1.2 Messidor
    	1.3 RIM-ONE
    	1.4 DRION-DB
    2 Multiscale sequential convolutional neural networks for simultaneous detection of fovea and optic disc
    3 黄斑分割
    4 Post-Process Methods
    5 Patched Based Attention Unet Model

# 1 Optic Disc 数据集 #

视盘（optic disc）：全称视神经盘，有时候也被成为视神经乳头（optic nerve head）。在普通的彩色眼底相机中，一般最亮的区域就是视盘的位置。

## 1.1 ORIGA-650 ##

这批数据集归属于新加坡国家眼科中心，主要包含650张彩色眼底图像，每张图像都有视盘和视杯的分割标注，同时还有是否患有青光眼的诊断标注。拥有这批数据的IMED团队，也是目前国内最大的眼科医疗图像组。 ORIGA-650分为两个子集 set A for training 和 set B for testing 每个子集包含了325张图像。

## 1.2 Messidor ##

Messidor数据集原本是用来做糖尿病视网膜（diabetic retinopathy， DR）病变检测的，只有糖网的分级标注。后来国外的一个课题组又重新手工标定了视盘的边界，因此目前大家也同样在Messidor数据上做视盘的定位和分割。  
The performance of our model on Messidor-dataset：  
![在这里插入图片描述][20210503111536812.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]

## 1.3 RIM-ONE ##

RIM-ONE一共发布了三个子数据集（RIM-ONE-R1，R2，R3），他们的数量分别是169，455和159张。

## 1.4 DRION-DB ##

DRION-DB 做的人特别少，但是这批数据集也有111张图像。大家也可以做一下，就当作一种data augmentation了吧。

PS：ISBI2019也有一个challenge，是关于AMD的分类，其中一个sub-task也是optic disc segmentation，

![在这里插入图片描述][20210503111635796.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
![在这里插入图片描述][20210503111651703.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
![在这里插入图片描述][20210503111710905.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]

# 2 Multiscale sequential convolutional neural networks for simultaneous detection of fovea and optic disc #

Deep Learning的方法做视盘和黄斑中心凹的检测
    基于的数据库MESSIDOR和Kaggle
    图片预处理时，为了避免颜色引入的额外复杂度，都先预处理成灰度图像，然后再做图像增强

![在这里插入图片描述][20210503111823869.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
分别是灰度图和数据增强后的图像。  
网络结构中，先回归出两个矩形区域，分别是视盘和黄斑的矩形区域，然后分别回归出中心。  
![在这里插入图片描述][20210503111919886.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
网络采用比较简单的形式，降采样的程度也不深，只做了两层降采样，这也是可以理解的，毕竟在眼底图像中，黄斑和视盘的范围还是很大的。当然文章感觉也不是那么严禁，要想把4和5对上还挺难的。不过毕竟是提供一种思路，可以自己尝试下  
![在这里插入图片描述][20210503111941932.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]

# 3 黄斑分割 #

![在这里插入图片描述][20210503112102124.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
![在这里插入图片描述][202105031124089.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
ROSE：视网膜OCTA血管分割数据集

![在这里插入图片描述][20210503112510536.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]  
(A) 眼底彩照，绿框为黄斑区域；（B-D）不同深度的OCTA图像。

# 4 Post-Process Methods #

When directly use unet model, we often get some error predictions. So I use a post-process algorithm:
    
    predicted area can't be to small.
    minimum bounding rectangle's height/width or width/height should be in 0.45~2.5
    lefted area is the final output. The problem of this algorithm is that the parameters not self-adjusting, so you have to change them if input image is larger or smaller than before.

# 5 Patched Based Attention Unet Model #

I use a modified Attention Unet which input of model is 128x128pix image patches. 
     When sampling the patches,
      I focus the algorithm get samples around optic disc.
       The patches is like that:

![在这里插入图片描述][2021050311284689.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]

[20210525100706727.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/149e4b50fd4d4337a3a3c0f6461cd520.png
[20210525101156408.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/e1ba97126f9c4a0fa41f735331632a4f.png
[20210503111536812.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/01594c4ee00a40439ecff304571a1a93.png
[20210503111635796.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/50618794f7ba4c0a83b0f1aa7c9926f9.png
[20210503111651703.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/5a52cd87882b4ebfb85b27c51b67f762.png
[20210503111710905.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/198d709196004fe680a724f3a4fe3c39.png
[20210503111823869.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/eab2b5b5640a498089f1ad939b7b9b75.png
[20210503111919886.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/553dd2c2b0334896927e4512aadad548.png
[20210503111941932.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/e83e09e77b97461a9222dbb7eabe3d09.png
[20210503112102124.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/4acbe8e6813e4634beeb22d9c178bbed.png
[202105031124089.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/ed7c9b150f2c4b6d80f5800395b26ca2.png
[20210503112510536.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/b133a86976cc4b9b97020922ba8ef100.png
[2021050311284689.png_x-oss-process_image_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MzgzODc4NQ_size_16_color_FFFFFF_t_70]: /images/20221021/4bbdafe05af047749cf6f8c8783a80c0.png