Abstract:Medical image segmentation is a crucial preliminary step for a number of downstream diagnosis tasks. As deep convolutional neural networks successfully promote the development of computer vision, it is possible to make medical image segmentation a semi-automatic procedure by applying deep convolutional neural networks to finding the contours of regions of interest that are then revised by radiologists. However, supervised learning necessitates large annotated data, which are difficult to acquire especially for medical images. Self-supervised learning is able to take advantage of unlabeled data and provide good initialization to be finetuned for downstream tasks with limited annotations. Considering that most self-supervised learning especially contrastive learning methods are tailored to natural image classification and entail expensive GPU resources, we propose a novel and simple pretext-based self-supervised learning method that exploits the value of positional information in volumetric medical images. Specifically, we regard spatial coordinates as pseudo labels and pretrain the model by predicting positions of randomly sampled 2D slices in volumetric medical images. Experiments on four semantic segmentation datasets demonstrate the superiority of our method over other self-supervised learning methods in both semisupervised learning and transfer learning settings. Codes are available at https://github.com/alienzyj/PPos.

Key words:self-supervised learning, medical image analysis, semantic segmentation

摘要：医学图像分割是许多下游诊断任务中的关键步骤。随着深度卷积神经网络极大地促进了计算机视觉的发展，半自动化的医学图像分割方法逐渐成熟，即通过应用深度卷积神经网络来检测感兴趣区域，然后由放射科医生进行修改。然而，有监督学习需要大量的人工标注，这些标注数据很难获得，特别是在医学图像领域。自监督学习能够利用无标签数据，为模型提供良好的初始化参数，然后在带标签数据量有限的下游任务上进行微调。考虑到大多数自监督学习特别是对比学习主要应用于自然图像领域，并且在预训练过程中需要昂贵的GPU资源，我们提出了一种新颖而简单的基于辅助任务的自监督学习方法，该方法利用了三维医学图像中的位置信息。具体来说，我们将二维切片在三维坐标系中的纵坐标作为伪标签，在预训练阶段以模型预测该标签作为辅助任务。我们在四个语义分割数据集上证明了本文的方法在医学图像分割任务中优于其他自监督学习方法。代码已在https://github.com/alienzyj/PPos 公开。

关键词:自监督学习，医学图像分析，语义分割