Characteristic patch-based deep and handcrafted feature learning for red lesion segmentation in fundus images

The appearance of red lesions on the retina can be an early indicator of diabetic retinopathy (DR, hereafter). Their timely detection can help mitigate the effects of DR and eventually save the vision. This paper proposes a DR-specific red lesion extraction using color fundus images. In the pre-processing stage, we suggest the removal of retinal blood vessels to enhance the red abnormalities and eventually minimize the false positives. To improve feature learning, we propose an integration of deep and handcrafted features. Here, we propose a set of five lesion-specific, handcrafted features and concatenate them with the U-net-based deep features. Moreover, to enhance the efficacy of the training process, we propose a novel 'characteristic patch-based training' to target the red lesions explicitly. In the paper, we also perform an ablation study to prove the significance of the handcrafted features and the characteristic training in improving the baseline model's performance. To the best of our knowledge, the proposed algorithm has out-performed the other state-of-the-art algorithms on IDRiD and DDR datasets by approximately 7% and 10%, respectively, in terms of area under the precision-recall curve.

Automated summary

This paper proposes a DR-specific red lesion extraction method using color fundus images. It integrates deep and handcrafted features and utilizes characteristic patch-based training to explicitly target the red lesions. Experimental results show that the proposed algorithm out-performs other state-of-the-art algorithms on two datasets.