| Title:
|
Improved microaneurysm detection using convolutional neural network
|
| Author(s):
|
Noushin Eftekhari, Hamidreza Pourreza
|
| Presentation Type:
|
Poster
|
| Subject:
|
Retina and Retinal Cell Biology
|
| Others:
|
|
|
Presenting Author:
|
|
| Name:
|
Noushin Eftekhkari
|
| Affiliation :(optional)
|
|
| E mail:
|
noushin.eftekhari@gmail.com
|
| Phone:
|
|
|
Mobile: |
09359161878
|
|
Abstract (Max 200 words)
|
|
| Purpose:
|
Diabetic retinopathy is an important reason of blindness in the world. Microaneurysm is the earliest clinical sign of diabetic retinopathy which if diagnosed early, the disease can be cured successfully. This paper proposed a new method based on convolution neural network for the automatic detection of microaneurysms (MAs) in the color retinal images. Convolutional neural networks (CNNs) are widely used in computer vision application. CNNs are based on convolution operations that are applied to the image at multiple hierarchical layers. The proposed neural network was trained to classify the image pixels, using patches which are centered in each pixels. The proposed method has been tested in the Retinopathy Online Challenge images.
|
| Methods:
|
In this paper, we propose a three-step algorithm for detection of MAs using convolutional neural network. Color retinal images often have non-uniform illumination. So, microaneurysms are hardly visible in regions with low brightness and poor contrast. Therefore, in the first step as a pre-processing, a color normalization method is applied to improve the image quality. Afterwards, in the second step, we train the convolutional neural network. The network output is a probability map which indicates the probability of microaneurysm existence. But the probability map have sever false positive rate. Therefore, in the final step, a simple post-processing method is applied to the network output. In this step, a kernel disk with radius five is convolved to the output and as a result, the regions with low probability values are removed.
|
| Results:
|
The proposed microaneurysm detection algorithm is evaluated on ROC database and the sensitivity result is equal to 70%. In the future, we are going to improve the algorithm so that better results can be achieved.
|
| Conclusion:
|
This paper introduces a convolutional neural network that can automatically detect microaneurysms in the retinal images and help to diagnose DR in retinal patient images. The proposed network consists of 11 layers. The evaluation results show the high performance for MA detection. Therefore, it is useful for diabetic retinopathy screening systems.
|
| Attachment:
|
5040sample.pptx
|