Title: Blind Inpainting with Object-aware Discrimination for Artificial Marker Removal

Abstract: Medical images often contain artificial markers added by doctors, which can negatively affect the accuracy of AI-based diagnosis. To address this issue and recover the missing visual contents, inpainting techniques are highly needed. However, existing inpainting methods require manual mask input, limiting their application scenarios. In this paper, we introduce a novel blind inpainting method that automatically completes visual contents without specifying masks for target areas in an image. Our proposed model includes a mask-free reconstruction network and an object-aware discriminator. The reconstruction network consists of two branches that predict the corrupted regions with artificial markers and simultaneously recover the missing visual contents. The object-aware discriminator relies on the powerful recognition capabilities of the dense object detector to ensure that the markers of reconstructed images cannot be detected in any local regions. As a result, the reconstructed image can be close to the clean one as much as possible. Our proposed method is evaluated on different medical image datasets, covering multiple imaging modalities such as ultrasound (US), magnetic resonance imaging (MRI), and electron microscopy (EM), demonstrating that our method is effective and robust against various unknown missing region patterns.