基于多任务学习的弱监督皮肤图像分割

Weakly Supervised Skin Image Segmentation Based on Multi Task Learning

为了降低网络受到数据集标注的限制,提出了一种基于多任务学习的弱监督医学图像语义分割模型,仅使用图像级真实标签实现了优异的医学图像分割性能.该模型利用图像分类任务和显著性检测任务辅助语义分割任务使之达到更好的分割效果,图像分类分支使用重建类激活图的方法降低类激活图过激活和欠激活的概率,同时设计了跨任务语义挖掘模块学习显著性检测任务和语义分割任务间的相似性用于优化特征图.融合类激活图和显著图生成像素级伪标签,并在训练过程中迭代优化伪标签,提升网络的分割性能.实验证明提出的方法在ISBI2016和ISIC2017皮肤图像数据集的平均交并比指标分别为68.24%和60.92%,远高于其他先进的弱监督语义分割算法.

It is well known that applying deep learning techniques for medical image segmentation requires the use of a large number of datasets with pixel-level annotations,however,pixel-level annotation of medical images is expensive.In order to reduce the limitation of the network by dataset labeling,a weakly supervised medical image semantic segmentation model based on multi-task learning is proposed,which only uses image-level ground-truth labels to achieve excellent medical image segmentation performance.The image classification branch uses the method of reconstructing the class activation map to reduce the probability of over-activation and under-activation of the class activation map,and designs the cross-task semantic mining module to learn the similarity between the saliency detection task and the semantic segmentation task to optimize the feature map.Finally,the fusion class activation map and saliency map generate pixel-level pseudo-labels,and iteratively optimize the pseudo-labels during the training process to improve the segmentation performance of the network.The experiments results show that the mean intersection over union of the proposed method in the skin image datasets of ISBI2016 and ISIC2017 is 68.24%and 60.92%,respectively,which is much higher than that of other advanced weakly supervised semantic segmentation algorithms.