Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
, 18 (1), 288

Fundus Images Analysis Using Deep Features for Detection of Exudates, Hemorrhages and Microaneurysms

Affiliations

Fundus Images Analysis Using Deep Features for Detection of Exudates, Hemorrhages and Microaneurysms

Parham Khojasteh et al. BMC Ophthalmol.

Abstract

Background: Convolution neural networks have been considered for automatic analysis of fundus images to detect signs of diabetic retinopathy but suffer from low sensitivity.

Methods: This study has proposed an alternate method using probabilistic output from Convolution neural network to automatically and simultaneously detect exudates, hemorrhages and microaneurysms. The method was evaluated using two approaches: patch and image-based analysis of the fundus images on two public databases: DIARETDB1 and e-Ophtha. The novelty of the proposed method is that the images were analyzed using probability maps generated by score values of the softmax layer instead of the use of the binary output.

Results: The sensitivity of the proposed approach was 0.96, 0.84 and 0.85 for detection of exudates, hemorrhages and microaneurysms, respectively when considering patch-based analysis. The results show overall accuracy for DIARETDB1 was 97.3% and 86.6% for e-Ophtha. The error rate for image-based analysis was also significantly reduced when compared with other works.

Conclusion: The proposed method provides the framework for convolution neural network-based analysis of fundus images to identify exudates, hemorrhages, and microaneurysms. It obtained accuracy and sensitivity which were significantly better than the reported studies and makes it suitable for automatic diabetic retinopathy signs detection.

Keywords: Convolutional neural networks; Deep learning; Diabetic retinopathy; Fundus image analysis; Image processing.

Conflict of interest statement

Ethics approval and consent to participate

Not applicable.

Competing interests

Not applicable.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Figures

Fig. 1
Fig. 1
Example of Retina Images containing three DR sings. This image shows an entire retina image with haemorrhage, microaneurysm and exudate labled by graders, and which was then cropped to illustrate individual patch
Fig. 2
Fig. 2
Overview of the proposed framework contains two main phases: 1) patch-based and 2) image-based analysis. The patch-based section corresponds to training and testing a CNN model to discriminate between the different DR signs. Image-based analysis of the entire image generates probability maps for each sign
Fig. 3
Fig. 3
Applying the image enhancement technique on an example retina image. (a) Original retina image; (b) After image enhancement. This shows that some new lesions can be singularized by image enhancement shown by yellow annotations)
Fig. 4
Fig. 4
Hierarchical architecture of the proposed CNN. I: input image, C: convolutional layer, FM: feature map, MP: max pooling, NM: Normalization layer, FC: fully-connected layer
Fig. 5
Fig. 5
Process of generating three probability maps corresponding to exudate, hemorrhage and microaneurysm from a retina image. By taking a patch of size S × S centered around pixel (xi, yi), each patch is fed to the trained CNN that determines the membership probabilities at location (xi, yi) for the three pathological signs: i.e. exudate, hemorrhage and microaneurysm (shown by PE,xi,yi, PH,xi,yi and PM,xi,yi)
Fig. 6
Fig. 6
Patch examples corresponding to the four classes; (a) exudate. b hemorrhage. c microaneurysm. d no-sign
Fig. 7
Fig. 7
Relationship between number of training epochs with accuracy over 100 epochs. It is observed that the accuracy saturated after 43th epoch to 90% and hence was selected as the maximum number of training epochs
Fig. 8
Fig. 8
Three probability maps were generated from an example retina image: (a) original retina image; (b) Exudate probability map; (c) Hemorrhage probability map; (d) Microaneurysm probability map. Colorbar shows the severity level of a pixel belong to the sign that is ranging between 0 to 1 corresponding to blue to red color
Fig. 9
Fig. 9
Three examples of pathological signs before and after post-processing. a Original image. b Probability map corresponding to the sign. c Image output after post-processing
Fig. 10
Fig. 10
ROC curve corresponding classification of the four classes (exudate, hemorrhage, microaneurysm and no-sign)
Fig. 11
Fig. 11
Performance of proposed framework for the sign detections using two databases (DIARETDB1 and e-Ophtha) compared to the method with binary outputs of the network
Fig. 12
Fig. 12
Segmentation output image of the example retina image. a Manually annotated images that exudate, hemorrhage, and microaneurysm signs marked by blue, green and pink color, respectively. b Segmented output by the proposed algorithm

Similar articles

See all similar articles

Cited by 3 PubMed Central articles

References

    1. Leontidis G. A new unified framework for the early detection of the progression to diabetic retinopathy from fundus images. Comput Biol Med. 2017;90:98–115. doi: 10.1016/j.compbiomed.2017.09.008. - DOI - PubMed
    1. Mookiah MRK, Acharya UR, Chua CK, Lim CM, Ng EYK, Laude A. Computer-aided diagnosis of diabetic retinopathy: a review. Comput Biol Med. 2013;43(12):2136–2155. doi: 10.1016/j.compbiomed.2013.10.007. - DOI - PubMed
    1. Kaur J, Mittal D. A generalized method for the segmentation of exudates from pathological retinal fundus images. Biocybernetics Biomed Eng. 2018;38(1):27–53. doi: 10.1016/j.bbe.2017.10.003. - DOI
    1. Shaw JE, Sicree RA, Zimmet PZ. Global estimates of the prevalence of diabetes for 2010 and 2030. Diabetes Res Clin Pract. 2010;87(1):4–14. doi: 10.1016/j.diabres.2009.10.007. - DOI - PubMed
    1. Mohamed Q, Gillies MC, Wong TY. Management of diabetic retinopathy: a systematic review. JAMA. 2007;298(8):902–916. doi: 10.1001/jama.298.8.902. - DOI - PubMed

MeSH terms

Feedback