Purpose :
Ultra-widefield (UWF) scanning laser ophthalmoscopy (SLO) imaging is a promising modality for diabetic retinopathy (DR). Manual segmentation of lesions on UWF images can be labor-intensive and error-prone. Therefore, there is a need for an automated lesion segmentation tool for UWF images.
Methods :
Our novel UWF retinal image analysis framework comprises of the following steps: (i) image enhancement that normalizes intensity and other variations, (ii) multi-scale morphological filtering based interest region detection that selects < 1% of pixels for further analysis, (iii) state-of-the-art deep learning techniques for estimating lesion likelihood.
225 UWF images of diabetic patients captured using Optos UWF cameras were annotated for hemorrhages, exudates, and cotton-wool spots by experts at Doheny Eye Institute. 60 images were used for training the system, and the remaining 165 images were analyzed to produce lesion likelihood maps as shown in Figure 1
Results :
On the independent dataset of 165 UWF images the preliminary lesion level performance statistics are as follows: (i) Hemorrhages: Sensitivity of 35.9% with false positive rate of 2.5%. (ii) Exudates: Sensitivity of 34.4% with false positive rate of 2.3%. (ii) Cottonwool spots: Sensitivity of 19.6% with false positive rate of 1.8%.
Conclusions :
The preliminary results on this independent and challenging dataset are promising. Such an automated lesion annotation tool will be valuable in quantitative analysis of UWF images and could enable accelerated research on DR progression.
Study Link:
- Sandeep Bhat, Christian Siagian, Chaithanya Ramachandra, Malavika Bhaskaranand, Connie Martin Sears, Muneeswar Gupta Nittala, Srinivas R. Sadda, and Kaushal Solanki. “Clinical Validation of Diabetic Retinopathy Lesion Segmentation in Ultra-Widefield Images.” Investigative Ophthalmology & Visual Science 59, no. 9 (July 13, 2018): 1666–1666.