Small Object Segmentation In 3D CT Images

Automatic 3D computed tomography(CT)medical image segmentation plays a crucial role in many clinical applications,such as disease diagnosis,surgical planning,evaluation of efficacy,etc.Recently,deep learning methods based on convolutional neural networks(CNN)have achieved remarkable performance in medical image segmentation.However,it remains challenging for small target segmentation as the target occupies a very small volume in an entire 3D CT image,and exhibits a large variation in its appearance(e.g.irregular shape,low-contrast texture,and blurred boundary).To alleviate this problem,recent state-of-the-art employs a two-stage network,which typically consists of two sub-networks,i.e.,the first stage subnet to eliminate irrelated background by roughly localizing the target,followed by the second stage subnet to obtain precise pixel-level predictions.Despite great success,existing two-stage networks often treat the two stages separately,ignoring the inherent feature correlation and error tolerance between each other.In this paper,we introduce a cross-stage communication path in two-stage networks for small target segmentation,which 1)make the second-stage sub-network inherit well-learned contextual features from the first stage,alleviating the complexity of pixel-level prediction by focusing on only learning complementary features,and 2)force the second-stage sub-network to perceive the final segmentation errors,yielding more segmentation-friendly regions of target for the subsequent stage.Besides,inspired by the fact that radiologists typically delineate the small targets with blurry boundaries in one slice according to the adjacent slices,fully utilizing the inter-slice context that objects in adjacent slices of a volume usually have intrinsic correlations in terms of shape and location,we further propose an inter-slice context aggregation method with two additional tasks to capture the correlation between adjacent slices from the perspectives of consistency and difference.Extensive experimental results on two datasets demonstrate the superior performance of our proposed methods to state-of-the-art methods.Our method achieves 84.80% dice similarity coefficient(DSC)in the public abdominal CT pancreas segmentation dataset,and83.95% in locally collected lung CT pneumonia segmentation dataset,respectively.