基于委员会查询和自步多样性学习的医学图像分割

Self-paced learning with diversity for medical image segmentationby using query-by-committee

由于拥有像素级标记的医学图像数量非常少,制约了卷积神经网络在医学图像分割任务上的应用,因此,该文提出了一种基于委员会查询的自步多样性学习算法,在训练数据有限的情况下提升医学图像分割模型的性能。该文所提算法结合了基于委员会查询的数据选择方法,实现动态地从易到难选择样本,对模型进行训练。同时,该算法通过应用仿射传播聚类,保证了数据选择的多样性,提升了图像分割模型的性能。为了验证所提算法框架的有效性,分别在3类医学图像分割任务的5个不同数据集任务上进行了实验,实验结果表明,该文所提算法可以显著提升分割性能。在使用相同数据的训练的情况下,相比于全监督学习,使用该文算法可以得到更高的Dice评估指标、表面距离和平均交并比值。

Because the number of pixel-wise labeled medical images are extremely small,which prevents the application of convolutional neural network(CNN)in medical image segmentation tasks,a query-by-committee based self-paced learning with diversity(SPLD)framework is proposed to boost the performance of medical image segmentation with limited data.The proposed SPLD algorithm combines the data selection method based the query-by-committee to realize the dynamic selection of samples from easy to difficult and train the model.Meanwhile,by applying the affine propagation clustering,the proposed algorithm guarantees the diversity of data and the performance of image segmentation model is further enhanced.To verify the effectiveness of the proposed QBC based SPLD framework,we conducted experiments on three medical image segmentation tasks with five different datasets.The experimental results show that the proposed algorithm can significantly improve the segmentation performance.With the same dataset,the proposed SPLD could significantly improve the segmentation performance and achieve a higher Dice score,surface distance and mean Intersection over Union(mIoU)than fully supervised learning.