A novel color space of fundus images for automatic exudates detection. (March 2019)

• Record Type: Journal Article
• Authors: Khojasteh, Parham; Aliahmad, Behzad; Kumar, Dinesh Kant
• Is Part Of: Biomedical signal processing and control. Volume 49(2019)
• Page Start: 240
• Page End: 249
• Publication Date: 2019-03
• Subjects: Exudate detection -- Convolutional neural networks -- Color space representation -- Machine learning -- Retinal image analysis -- Deep learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2018.12.004
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Accurate and automatic detection of exudates in fundus images. Assessing performance of the different color spaces of fundus images using a convolutional neural network. Proposing a novel space by combination of three channels. The most reliable and repeatable results were obtained by the proposed space compared to previous studies. Abstract: This paper has compared the performance of different color spaces of fundus images for automatic detection of exudates. A convolutional neural network was employed to assess the performances of different color spaces generated by orthogonal transformation of the original colors in red/green/blue (RGB) space. Experiments were conducted on two publicly available databases: (1) DIARETDB1 and (2) e-Ophtha. Based on the experimental results, this study has proposed a new color space of fundus images with three channels: (i) second eigenchannel of the RGB space, (ii) hue and (iii) saturation channels of Hue/Saturation and Intensity (HSI) space. This achieved an accuracy, sensitivity and specificity of 98.2%, 0.99 and 0.98, respectively. Twenty times 20-fold cross validation technique confirmed that proposed color space obtained higher replicability compared with conventional color spaces.