Joint disease classification and lesion segmentation via one-stage attention-based convolutional neural network in OCT images

Abstract Optical coherence tomography (OCT) is a useful tool for the diagnosis of macular diseases. It is necessary to identify macular diseases and segment lesion areas for assisting ophthalmologist in clinical diagnosis. Deep learning-based methods have been proposed to solve this issue. However, most of them used segmentation task to guide classification task and two tasks are optimized independently so that the information learned by the classification task cannot be transferred to the segmentation task. To overcome the problems, we propose a one-stage attention-based method for retinal OCT image classification and segmentation in bounding box level supervision. Specifically, the classification network is used to generate heatmap by Gradient-weighted Class Activation Mapping and introduces the proposed attention block. Transformation consistency is used for encouraging the predicted heatmap to be consistent for the same input after the image transformation. The grayscale feature map (GFM) is proposed based on grayscale prior information. GFM is applied to dynamically constrain the heatmap for further gaining stronger discriminative information and develop attention blocks as an attention constraint. Experimental results on two datasets verify the efficiency of the proposed method.