基于深度学习分割模型的脑出血CT图像自动分割研究

Research on Automatic Segmentation of CT Images of Intracerebral Hemorrhage Based on Deep Learning Segmentation Model

目的实现基于人工智能深度学习方法的脑出血患者CT图像血肿区域自动分割,并评估后处理算法对分割结果的优化效果。方法回顾性分析吉林大学第一医院2018年4月至2020年8月脑出血患者的影像资料,按照纳排标准纳入416例患者的CT图像进行研究,按照比例7∶3随机分为训练集291例和测试集125例。将患者CT图像进行图像预处理、摆正、去骨后,采用本研究提出的深度学习分割网络ADUNET进行训练,实现血肿区域的自动分割。最后使用后处理算法对分割结果进行进一步优化,并通过Dice系数、Hausdorff-Distance(HD)系数等评价指标进行对比分析。结果对比其他两种主流分割网络,本研究提出的ADUNET网络在该数据集上得到了最好的分割结果(平均Dice系数为0.895,平均HD系数为11.62),并且验证了后处理算法可以进一步优化分割结果,提高分割精度(平均Dice系数为0.899、平均HD系数为11.33)。结论本研究提出的ADUNET分割网络与后处理算法可以实现基于CT图像的脑出血区域自动分割及优化,该方法可提高诊断效率、优化诊断流程,具有较高的临床应用价值。

Objective To achieve automatic segmentation of hematoma region in CT image of patients with intracerebral hemorrhage based on artificial intelligence deep learning method,and evaluate the optimization effect of post-processing algorithm on segmentation results.Methods The imaging data of patients with intracerebral hemorrhage in the First Hospital of Jilin University from April 2018 to August 2020 were retrospectively analyzed.416 patients were included in the study according to the inclusion criteria.They were randomly divided into 291 cases in the training set and 125 cases in the test set according to the ratio of 7∶3.After the CT images of patients were preprocessed,straightening and bone removal,an automatic segmentation network ADUNET proposed in this study was used for training to realize the automatic segmentation of hematoma area.Finally,the post-processing algorithm was used to further optimize the segmentation results,and the Dice coefficient,Hausdorff-Distance(HD)coefficient and other evaluation indicators were compared and analyzed.Results Compared with other two mainstream segmentation networks,ADUNET proposed in this design achieved the best segmentation results on this data set(average Dice was 0.895,average HD was 11.62),and verifies that the post-processing algorithm could further optimize the segmentation results and improve the segmentation accuracy(average Dice was 0.899,average HD was 11.33).Conclusion The ADUNET segmentation network and post-processing algorithm proposed in this study can realize the automatic segmentation and optimization of intracerebral hemorrhage area based on CT image.This method can improve the diagnosis efficiency and optimize the diagnosis process,and has high clinical application value.