Abstract
Purpose :
Artificial intelligence algorithms known as “deep-learners” have achieved great performance on a variety of automatic image recognition tasks in recent years. In this study, we investigate the use of a so called fully convolutional networks (FCNs) for image segmentation of specific retinal lesions indicative of diabetic retinopathy (DR) or diabetic macular edema (DME) in retinal fundus images.
Methods :
Using 45-degree field of view retinal fundus images (n=248) we trained two instances of the a pre-trained FCN-8s architecture for digital image segmentation of the specific DR or DME lesions; micro-aneurysms, hemorrhages and hard exudates. Separate valdidation images (n=5) and test images (n=5) used for tuning network parameters and evaluating network performance respectively were not used for training. Data augmentation was utilized to increase the number of training images and to create two distinct training sets of full sized images (n=5,456) and small image patches (n=463,760) used for training two different networks. Image patches not containing any lesions were discarded prior to training. Performance for both networks were evaluated on the test images (n=5) according to two different performance metrics. A hit and miss metric was used to evaluate the segmentation performance for individual lesion blobs and the f1-score was used to evaluate pixel-level segmentation performance for each lesion type.
Results :
Both networks performed reasonably well for detection of hard exudates (f1-scores > 0.72). With regards to hemorrhages and micro-aneurysms, the results indicate that the network trained on image patches performs better, as it correctly identifies more of these lesion types in the five test images. Likewise, recall increased from 0.08 to 0.45 for micro-aneurysms and 0.30 to 0.59 for hemorrhages when training on patches instead of full-sized images.
Conclusions :
The result of this study indicate that FCNs can be used for automatic extraction of DR and DME lesions in retinal fundus images. These results are encouraging with regards to the use of artificial intelligence and in particular deep-learning for automatic retinal fundus image analysis. Future studies should aim to collect more high quality data and the use of small image patches for training, as this seems to benefit performance.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.