共享浅层参数多任务学习的脑出血图像分割与分类

Intracerebral Hemorrhage Image Segmentation and Classification Based on Multi-task Learning of Shared Shallow Parameters

非增强CT扫描是急诊室诊断疑似脑出血的首选方法,医疗人员通常借助CT图像对疑似急性脑出血患者病灶部位进行手动分割,进而根据临床经验进行分类,这种人工诊断的方式对医师的经验要求较高,主观性较强,将分割和分类任务分开执行,不能充分利用两个任务间相关联的特征信息,时间成本高,增大了基于CT图像快速进行脑出血病灶部位分割及分类的难度。针对上述问题,文中提出了一种共享浅层参数多任务学习的脑出血图像分割及分类模型,一方面,根据不同任务学习的难易程度对损失函数的权值进行优化,另一方面,在多任务学习网络的浅层实现公有信息共享,深层提取不同任务的私有信息,获取更具代表性的特征,从而快速、准确地对脑出血患者的CT图像进行分割及分类。实验结果表明,共享浅层参数多任务学习网络生成的分割标注与真实标注有较好的视觉一致性。在最优权值下所有被试的平均Dice系数(DSC)为0.828,敏感度为0.842,特异度为0.985,阳性预测值(PPV)为0.838。共享浅层参数多任务学习网络分类的准确率、敏感度、特异度和AUC值分别为95.00%,90.48%,100.00%和0.982。与单任务深度学习、Y-Net以及借助分类辅助的多任务学习相比,该方法更加有效地利用了相关任务信息,同时通过调节损失函数权值,提升了出血病灶区域的分割和分类精度。

Non-enhanced CT scanning is the first choice for the diagnosis of suspected cerebral hemorrhage in the emergency room.Medical staffs usually use CT images to manually segment the lesions of patients with suspected acute cerebral hemorrhage,and then classify them based on clinical experience.This method of manual diagnosis requires the physician’s experience and is highly subjective.Moreover,the segmentation and classification tasks are performed separately,and the characteristic information associated between the two tasks cannot be fully utilized,and the time cost is high,which increases the difficulty of quickly segmenting and classifying cerebral hemorrhage lesions based on CT images.In response to the above problems,the paper proposes a model for segmentation and classification of cerebral hemorrhage images based on multi-task learning.On the one hand,the weight of the loss function is optimized according to the difficulty of learning different tasks.On the other hand,public information sharing is realized in the shallow layer of the multi-task learning network,and private information of different tasks is extracted deeply to obtain more representative features,so as to quickly and accurately segment and classify the CT images of patients with cerebral hemorrhage.The experimental results show that the segmentation annotations generated by the multi-task learning network have good visual consistency with the real annotations.Under the optimal weight,the average Dice coefficient(DSC)of all subjects is 0.828,the sensitivity is 0.842,the specificity is 0.985,and the positive predictive value(PPV)is 0.838.The accuracy,sensitivity,specificity and AUC value of multi-task learning network classification are 95.00%,90.48%,100.00%and 0.982,respectively.Compared with single-task deep learning,Y-Net and multi task learning assisted by classification,this method makes more effective use of relevant task information,and at the same time improves the segmentation and classification accuracy of hemorrhage lesions by adjusting the weight of the loss function.