Pterygium-Net: a deep learning approach to pterygium detection and localization

• Published in: Multimedia tools and applications Vol. 78; no. 24; pp. 34563 - 34584
• Main Authors: Zulkifley, Mohd Asyraf, Abdani, Siti Raihanah, Zulkifley, Nuraisyah Hani
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2019; Springer Nature B.V
• Subjects: Algorithms; Artificial neural networks; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Deep learning; Failure rates; Image detection; Localization; Machine learning; Medical imaging; Medical personnel; Multimedia Information Systems; Neural networks; Rural areas; Sensitivity; Special Purpose and Application-Based Systems; Support vector machines; Training; Pterygium localization; Machine learning; Pterygium detection
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-019-08130-x

Automatic pterygium detection is an essential screening tool for health community service groups. It allows non-expert to perform screening process without the needs of big and expensive equipment, especially for the application in rural areas. Thence, patients who have been screened as positive pterygium will be referred to the certified medical personnel for further diagnosis and treatment. Current state-of-the-art algorithms for pterygium detection rely on basic machine learning approach such as artificial neural network and support vector machine, which have not yet achieved high detection sensitivity and specificity as required in standard medical practice. Hence, a deep learning approach based on fully convolutional neural networks is proposed to detect and localize the pterygium infected tissues automatically. The input image requirement for the developed system is low as any commercial mobile phone camera is sufficient. Moreover, the developed algorithm, which we refer as Pterygium-Net works well even if the eye image is captured under low lighting condition with pupil position is not at the center location. Pterygium-Net utilizes three layers of convolutional neural networks (CNN) and three layers of fully connected networks. Two steps are implemented to overcome lacks of training data by generating synthetic images and pre-training the CNN weights and biases in a different public dataset. As for pterygium localization, an additional step of box proposal based on edges information is used to generate possible regions of the pterygium infected tissues. Hanning window is also applied to the generated regions to give more weightage to the center area. Experimental results show that Pterygium-Net produces high average detection sensitivity and specificity of 0.95 and 0.983, respectively. As for pterygium tissues localization, the algorithm achieves 0.811 accuracy with a very low failure rate of 0.053. In the future, deeper networks can be implemented to further improve pterygium localization.