摘要
目的提出一种双任务自监督学习方法,该方法利用未经标注的医学影像数据提取通用视觉表征,以提升深度学习模型在颅脑CT扫描影像中(包括脑出血、鼻咽癌及常规检查影像)对脑组织的精准分割能力。方法结合图像补全与分类作为辅助手段,在70%图像信息被遮蔽的条件下,通过任务间轻量级的特征融合机制来优化预训练的输出表示。为评估该方法的有效性,研究采用平均Dice系数(mDice)和Hausdorff Distance 95%(HD95)作为评价指标,在经过本文自监督方法的预训练过程后,评估下游脑组织分割模型的性能表现。结果实验结果表明,本文方法有效提升了不同深度学习模型在分割任务中的性能。在单独监督学习环境下,ConvNeXt-V2-Base的mDice结果值最高,为0.8812,而本文分割方法在HD95指标方面表现最优,为7.95。与MAE方法相比,引入本文的自监督预训练策略后,本文分割方法表现出同比最佳的mDice(0.9589)和HD95(7.10)结果,且DenseNet121_Backbone的分割性能得到了显著提升,mDice指标上升至0.9512。结论本文的自监督预训练方法和编码器结构在CT医学影像数据中具备一定的表征学习优势,能够有效地提升下游脑组织分割任务的模型性能。
Objective To propose a dual-task self-supervised learning method,which uses unlabeled medical imaging data to extract general visual representations,aiming to enhance the precise segmentation capabilities of deep learning models in cranial CT scans(including cerebral hemorrhage,nasopharyngeal carcinoma,and routine scans).Methods Combined with image reconstruction and classification as auxiliary tasks,under the condition of 70%image information was occluded,a lightweight feature fusion mechanism between tasks was used to optimize the output representation of pre-training.To evaluate the effectiveness of this method,the average Dice coefficient(mDice)and Hausdorff Distance 95%(HD95)were used as evaluation indexes.After the pre-training process of the self-supervised method,the performance of the downstream brain tissue segmentation model was observed.Results Experimental results demonstrated that the proposed method effectively improved the performance of different deep learning models in segmentation tasks.In a standalone supervised learning environment,ConvNeXt-V2-Base achieved the highest mDice value of 0.8812,while the segmentation method performed optimally in terms of HD95 index of 7.95.Compared to the MAE method,the introduction of the self-supervised pre-training strategy resulted in the best mDice(0.9589)and HD95(7.10)outcomes for our segmentation method.Additionally,a significant improvement was observed in the segmentation performance of DenseNet121_Backbone,with the mDice metric increasing to 0.9512.Conclusion The self-supervised pre-training method and encoder model exhibit certain advantages in representation learning within CT imaging data,effectively enhancing the model performance for downstream brain tissue segmentation tasks.
作者
陈福军
孟令慧
王小红
CHEN Fujun;MENG Linghui;WANG Xiaohong(Liaoning Provincial Drug Evaluation and Inspection Center,Shenyang Liaoning 110003,China;School of Pharmacy,School of Intelligent Medicine,China Medical University,Shenyang Liaoning 110000,China;Department of Tissue Engineering Research,School of Intelligent Medicine,China Medical University,Shenyang Liaoning 110000,China)
出处
《中国医疗设备》
2025年第1期26-33,共8页
China Medical Devices
关键词
自监督学习
图像分割
表征学习
医学影像分析
深度学习
卷积神经网络
self-supervised learning
image segmentation
representative learning
medical imaging analysis
deep learning
convolutional neural network
