Weakly Semi-supervised object detection with point annotations in Retinal OCT images

Abstract

Optical coherence tomography (OCT) is a widely used ophthalmic imaging technique, and accurate detection of retinal biomarkers in OCT images can help physicians diagnose diseases. However, OCT images are not easy to obtain and are time-consuming and laborious. In addition, the size of biomarkers varies widely. Past deep learning-based methods can hardly solve the above problems well. Thus, to overcome the above challenges, we propose a weakly semi-supervised method called PO-Net for the detection of retinal biomarkers in OCT images. In the proposed method, we utilize a small set containing images with bounding box labels and a large set of weakly annotated images with only one point annotation per biomarker. The training of the net is composed of several steps. In the first step, we use the weakly annotated images to train a point-to-box regression network. In the second step, the point-annotated images are used to generate pseudo-bounding boxes. In the third step, the images with bounding box annotations and the generated images with pseudo-bounding box labels are used as inputs to the detection network. Furthermore, we propose a multi-scale feature fusion module to deal with the problem of biomarker appearance changes. The effectiveness of the proposed method is evaluated on a local dataset, and the state-of-the-art performance of our method is achieved in all datasets with different percentages of bounding box annotations.