摘要
针对传统心脏MRI图像分割人工成本高昂、效率低下的问题,提出弱监督学习框架和条件随机场的心脏MRI图像分割方法,通过双分支的网络结构,由局部交叉熵损失监督,从已有的涂鸦标注中学习;同时,利用两个解码器的输出增强模型训练,动态混合两个分支的输出生成伪标签,获得比涂鸦标注更准确的监督信号;结合涂鸦监督和伪标签监督,对分割网络进行训练;引入条件随机场对网络输出进行后处理,利用像素间关系提高分割结果的精确性。实验结果表明该模型在相近的标注代价条件下可获得比现有弱监督方法更优的分割性能。
Aiming at the high labor costs and low efficiency of traditional cardiac magnetic reso-nance imaging(MRI)segmentation,a cardiac MRI image segmentation method based on weakly super-vised learning and conditional random field is proposed.A dual branch structure is designed,which is supervised by partial cross-entropy loss to learn from existing scribble labels.In order to obtain more precise supervisory signals than scribble labeling,use the outputs of two decoders to enhance model training and dynamically mix the outputs of the two branches to generate pseudo labels.Then,train the segmentation network by combining scribble supervision and pseudo label supervision.Finally,intro-duce conditional random field for post-processing the output of the segmentation network,utilizing pixel relationships to enhance the accuracy of the segmentation results.The experiment results show that this method outperforms existing weakly supervised segmentation methods with similar annotation costs.
作者
张林
毕凯悦
李文宗
何俊彦
刘辉
ZHANG Lin;BI Kaiyue;LI Wenzong;HE Junyan;LIU Hui(School of Information and Control Engineering,China University of Mining and Technology,Xuzhou Jiangsu 21116,China)
出处
《佳木斯大学学报（自然科学版）》
CAS
2024年第7期1-5,共5页
Journal of Jiamusi University：Natural Science Edition
基金
国家自然科学基金面上项目(61971422)
江苏省学位与研究生教育教学改革课题(JGKT23_B037)。
关键词
医学图像分割
心脏磁共振图像
弱监督学习
涂鸦标注
探究性实验
medical image segmentation
cardiac magnetic resonance imaging
weakly supervised learning
scribble annotation
inquiry experiment
