Learning with Multitask Adversaries using Weakly Labelled Data for Semantic Segmentation in Retinal Images
Proceedings of The 2nd International Conference on Medical Imaging with Deep Learning, PMLR 102:414-426, 2019.
A prime challenge in building data driven inference models is the unavailability of statistically significant amount of labelled data. Since datasets are typically designed for a specific purpose, and accordingly are weakly labelled for only a single class instead of being exhaustively annotated. Despite there being multiple datasets which cumulatively represents a large corpus, their weak labelling poses challenge for direct use. In case of retinal images, they have inspired development of data driven learning based algorithms for segmenting anatomical landmarks like vessels and optic disc as well as pathologies like microaneurysms, hemorrhages, hard exudates and soft exudates. The aspiration is to learn to segment all such classes using only a single fully convolutional neural network (FCN), while the challenge being that there is no single training dataset with all classes annotated. We solve this problem by training a single network using separate weakly labelled datasets. Essentially we use an adversarial learning approach in addition to the classically employed objective of distortion loss minimization for semantic segmentation using FCN, where the objectives of discriminators are to learn to (a) predict which of the classes are actually present in the input fundus image, and (b) distinguish between manual annotations vs. segmented results for each of the classes. The first discriminator works to enforce the network to segment those classes which are present in the fundus image although may not have been annotated i.e. all retinal images have vessels while pathology datasets may not have annotated them in the dataset. The second discriminator contributes to making the segmentation result as realistic as possible. We experimentally demonstrate using weakly labelled datasets of DRIVE containing only annotations of vessels and IDRiD containing annotations for lesions and optic disc. Our method using a single FCN achieves competitive results over prior art for either vessel or optic disk or pathology segmentation on these datasets.