segmentation of optic cup Hong Kang,Kai Wang,Song Guo,Yingqi Gao, - - PowerPoint PPT Presentation

Pixel quantification for robust segmentation of optic cup

Hong Kang,Kai Wang,Song Guo,Yingqi Gao, Ning Li,Jinyuan Weng,Xiaoxing Li,Tao Li

1/10 上工医信 Beijing Shanggong HIS. Technology Co., Ltd

Pipeline

2/10

Pixel Quantification DeepLab v3+ Classification Fundus image binary Component Analysis binary Component Analysis

Figure 1. An overview of our proposed pipeline

Pipeline (Pixel Quantification)

3/10

Figure 2. Fundus images from training set and validation set

Pixel quantification is used to reduce the sensitivity of the segmentation model to color. 𝑦 is an image from training set, 𝑠 is a hyper-parameter.

Figure 3. Fundus image after pixel quantification

Eq (1)

Pipeline (Segmentation model)

4/10

Figure 4. An overview of DeepLab v3+

Patch: 900*900 Data augmentation: Rotation, flipping, scaling Hyper-parameters: learning rate 0.0005 max iteration 100000 momentum 0.9 weight decay 0.00004

Pipeline (Component Analysis)

5/10

False positive reduction:

• Binary probability map through a fixed

threshold (hyper-parameter)

• Find all connected component in the binary

map

• Reserve the largest connected component

Figure 5. (a) binary map (b) processed binary map

(a) (b)

Pipeline (Classification)

6/10

Glaucoma classification:

• Calculate vertical CDR (V-CDR)
• V-CDR is treated as the probability of getting

glaucoma

𝑤𝑑𝑒𝑠 = 𝑠𝑑 𝑠𝑒 Eq (2)

Figure 6. An example abot the calculation of vcdr.

Result

7/10

Ablation study on pixel quantification

• Pixel quantification can lead to 0.029

improvement in OC segmentation

• Pixel quantification can improve model

robustness to make it suitable for different fundus cameras.

Table 1. Segmentation results on validation set

Result

8/10

Segmentation Leaderboard

Result

9/10

Overall Leaderboard

Conclusion



Our solution

 Feature work

10/10  A pipeline for optic cup, optic disc segmentation and glaucoma classification  A robust segmentation model suitable for different fundus cameras  Interpretable deep learning model for glaucoma screening

Thanks! Q&A

上工医信 Beijing Shanggong HIS. Technology Co., Ltd