Performance evaluation of lightweight convolutional neural networks on retinal lesion segmentation

In addition to the recent development of deep learning-based, automatic detection systems for diabetic retinopathy (DR), efforts are being made to integrate those systems into mobile detection devices running on the edge requiring lightweight algorithms. Moreover, to enable clinical deployment it is important to enhance the transparency of the deep learning systems usually being black-box models and hence giving no insights into its reasoning. By providing precise segmentation masks for lesions being related to the severity of DR, a good intuition about the decision making of the diagnosing system can be given. Hence, to enable transparent mobile DR detection devices simultaneously segmenting disease-related lesions and running on the edge, lightweight models capable to produce fine-grained segmentation masks are required contradicting the generally high complexity of fully convolutional architectures used for image segmentation. In this paper, we evaluate both the runtime and segmentation performance of several lightweight fully convolutional networks for DR related lesion segmentation and assess its potential to extend mobile DR-grading systems for improved transparency. To this end, the U2-Net is downscaled to reduce the computational load by reducing feature size and applying depthwise separable convolutions and evaluated using deep model ensembling as well as single- and multi-task inference to improve performance and further reduce memory cost. Experimental results using the U2-Net-S† ensemble show good segmentation performance while maintaining a small memory footprint as well as reasonable inference speed and thus indicate a promising first step towards a holistic mobile diagnostic system providing both precise lesion segmentation and DR-grading.