基于多实例学习及阈值伪标签提取的CT影像颅内出血分割

Segmentation of Intracranial Hemorrhage in CT Images Based on Multi-Instance Learning and Thresholding Pseudo-Labels Extraction

颅内出血由颅内血管破裂引起,出血体积对治疗决策和预后分析具有重要的临床意义,而基于CT影像的血肿分割是体积测量的基础。全监督方法依赖于人工勾画的标签,十分耗时和繁琐,现有弱监督分割方法的鲁棒性差,容易受伪影干扰。为此,本研究提出了基于多实例学习的弱监督颅内出血分割网络MIL-ICH,由双分支结构组成。首先,由多实例学习解码器生成热图定位出血区域;然后,在热图基础上使用CT值阈值和像素自适应优化模块提取并优化伪标签,训练分割解码器;最后,两个分支同时训练,提高训练效率并且利用多分支协同作用进一步提升分割性能。在来自RSNA颅内出血数据集的200例CT扫描上的测试结果表明,MIL-ICH网络的Dice相似性系数和体积相似度分别达到了0.822和0.896,本网络测量的出血量与实际出血量的相关性优于临床常用的多田公式估测法。所提出的方法能够提高颅内出血弱监督分割性能,有助于为临床提供出血体积测量和预后评价的依据。

Intracranial hemorrhage is the bleeding caused by the rupture of intracranial blood vessels,and the volume of the hematoma is clinically important for treatment decision and prognosis analysis.The segmentation of the hematoma based on CT images is the basis of the volume measurement.Fully supervised methods rely on manually outlined labels,which are time-consuming and laborious,while existing weakly supervised segmentation methods have poor robustness and are prone to be affected by artifacts.To this end,this study proposed MIL-ICH,a multi-instance learning based weakly supervised network for intracranial hemorrhage segmentation.The network is composed of a two-branch structure.First,the multi-instance learning decoder generated heatmap to locate the hemorrhage area.Then,based on the heatmap,the pseudo-labels were extracted and optimized by CT value thresholding and pixel-adaptive refinement module to train the segmentation decoder.Finally,the two branches were trained simultaneously to improve training efficiency and leverage the multi-branch collaboration to further improve segmentation performance.The test results on 200 CT scans from the RSNA intracranial hemorrhage dataset showed that the Dice similarity coefficient and volume similarity of MIL-ICH reached 0.822 and 0.896,respectively.The correlation of the hematoma volume measured by this network with the actual hematoma volume is better than the ABC/2 estimation method commonly used in clinical practice.In conclusion,the method proposed in this work can improve the performance of weakly supervised segmentation of intracranial hemorrhage and benefit the volume measurement and prognostic evaluation for clinical purposes.