精准的医学图像自动分割是临床影像学诊断和影像三维重建的重要基础。但医学图像数据的目标对象间对比度差异小、受器官运动影响大，加之标注样本规模小，因此在小样本下建立高性能的医学分割模型仍是目前的难点问题。针对主流原型学习小样本分割网络对医学图像边界分割性能差的问题，提出一种迭代边界优化的小样本分割网络(Iterative boun-dary refinement based few-shot segmentation network，IBR-FSS-Net)。以双分支原型学习的小样本分割框架为基础，引入类别注意力机制和密集比较模块(Dense comparison module，DCM)，对粗分割掩码进行迭代优化，引导分割模型在多次迭代学习过程中关注边界，从而提升边界分割精度。为进一步克服医学图像训练样本少且多样性不足问题，使用超像素方法生成伪标签，扩充训练数据以提升模型泛化性。在ABD-MR和ABD-CT医学图像分割公共数据集上进行实验，与现有多种先进的医学图像小样本分割方法进行对比分析和消融实验。实验结果表明，该方法有效提升了未见医学类别的分割性能。
A Few-shot Medical Image Segmentation Network With Iterative Boundary Refinement
Accurate automatic segmentation of medical images is an important basis for clinical imaging diagnosis and 3D image reconstruction.However,medical image data has small contrast differences between target objects,is greatly affected by organ movement,and the scale of labeled samples is small.Therefore,it is still a difficult prob-lem to establish a high-performance medical segmentation model under few samples.In view of the poor perform-ance of the mainstream prototype learning few-shot segmentation network for medical image boundary segmenta-tion,an iterative boundary refinement based few-shot segmentation network(IBR-FSS-Net)is proposed.Based on the few-shot segmentation framework of dual-branch prototype learning,the category attention mechanism and dense comparison module(DCM)are introduced to iteratively refine the coarse segmentation mask,and guide the segmentation model to focus on the boundary during multiple iterative learning processes,thereby improving the boundary segmentation accuracy.In order to further overcome the problem of few training samples and insufficient diversity of medical images,this paper uses the super-pixel method to generate pseudo-labels and expand the train-ing data to improve the generalization of the model.Experiments on the mainstream ABD-MR and ABD-CT med-ical image segmentation public datasets are done,we conduct extensive comparative analysis and ablation experi-ments with various existing advanced medical image few-shot segmentation methods.The results show that our method effectively improves the segmentation performance of unseen medical categories.
Medical image segmentationfew-shot learningattention mechanismboundary refinement
万方数据
维普
