Deep learning is increasingly used in every medical imaging segmentation task, but detection of lesions from eye fundus images (EFI) poses many difficult challenges related to sizes, similarity with other lesions and structures, low contrasts, variant conformations. During training, the loss function directs backpropagation learning in the deep convolutional neural networks (DCNN) that are used. It is therefore a fundamental function to the optimization procedure. There exist alternative formulations, such as cross entropy, jaccard and dice. But does the choice of loss influence quality decisively, in the difficult context of EFI lesions? And what about the network architecture? As part of our effort to improve the approaches, we evaluate alternative loss functions, also alternative architectures. We show that the choice of a convenient architecture and loss function can double the quality detecting some of the small and difficult to detect lesions, but we also show that research is still required to find ways to improve the results further.
|