New paper at MICCAI 2019!

Our paper entitled “Multiclass segmentation as multitask learning for drusen segmentation in retinal optical coherence tomography” has been accepted for publication at MICCAI 2019!

Drusen segmentation in OCT is a challenging task. In general, it is posed either as a binary segmentation problem (drusen vs. everything else) or as a layer segmentation problem (by detecting the outer boundary of the RPE and the Bruch’s membrane). Each approach has its advantages and disadvantages: binary segmentation might be prone to false positives in areas that are not anatomically plausible, while layer segmentation might fail under the presence of large drusenoid deposits. In this paper we propose to take the best of each world by posing a multiclass segmentation problem in which we target both layers and drusen. Furthermore, instead of training a single, large capacity U-Net model with a multiclass cross entropy loss, we propose to benefit from multitask learning. To this end, we disentangle the multiclass problem as a series of binary segmentation tasks, and we assign to each of them its own decoder, while sharing a single encoder through skip connections. We also analyze if adding connections between the decoders for the layers with the decoder of the drusen segmentation task helps or not. In our experiments we observed better performance when no gradient flow is allowed through these extra connections.

If you have other problems with “sandwiched” classes, then maybe this idea can help you to boost your performance :)

This paper is part of Rhona Asgari’s PhD thesis, supervised by Hrvoje Bogunović at OPTIMA.