Weakly-supervised localization and classification of biomarkers in OCT images with integrated reconstruction and attention

Abstract

The retina biological markers play a crucial role in managing chronic eye conditions, and optical coherence tomography (OCT) is widely used for ophthalmic diseases. However, the established biomarkers have variable shapes and sizes, and pixel-level medical imaged labels are hard to obtain in practice, both making challenges to the localization and classification of them. Therefore, it is essential to locate and classify retinal biomarkers simultaneously under weak supervision. In this paper, we propose a novel method for weakly supervised localization and classification of biomarkers in OCT images, named LCG-Net, based on generative adversarial network (GAN) and attention, using image-level labels for training. The structure-texture information is used to improve the quality of the reconstructed healthy-alike image. And erasure strategy is introduced to ensure the attention maps cover biomarker regions wholly and obtain the biomarker class information. Besides, we fused residual mask in reconstruction and attention mask to locate biomarkers. Our results demonstrate that the proposed LCG-Net achieved higher DSC, Recall and IoU than other comparison methods on the local AB dataset and the public Cell dataset. The proposed method achieved averaged Dice 0.504 for seven disease-related biomarkers, and the Dice for healthy images is 0.922. Experiment results on two large datasets demonstrate the effectiveness of the proposed weakly supervised biomarker localization and classification framework.