多层特征融合的超声甲状腺结节分割方法
Ultrasound thyroid nodule segmentation method based on multi-layer feature fusion
张雅婷 1赵宸 1帅仁俊 1吴梦麟1
作者信息
- 1. 南京工业大学计算机科学与技术学院,江苏南京 211816
- 折叠
摘要
为精确地从超声图像中分割出甲状腺结节,提出一种包含Swin Transformer和卷积神经网络两个分支的多层特征融合分割方法,利用3个单向特征桥接单元(one-way feature bridging unit,OFU)桥接多层语义特征,并下采样特征图.实验采用来自斯坦福AIMI共享数据集的超声甲状腺结节图像用于训练、验证和测试.经过实验对比,验证了该模型在用时较短的情况下,相比其它模型取得了更好的分割效果.
Abstract
To accurately segment thyroid nodules from ultrasound images,a multi-layer feature fusion segmentation method con-taining two branches of Swin Transformer and convolutional neural network was proposed.Three one-way feature bridging units(OFU)were used to bridge multi-layer semantic features and downsample the feature map.Ultrasound thyroid nodule images from Stanford AIMI shared dataset were used for training,validation and testing.Through experimental comparison,it is veri-fied that the model achieves better segmentation effects than other models on the premise of less time consuming.
关键词
图像分割/甲状腺结节/特征融合/深度学习/特征提取/下采样/图像预处理Key words
image segmentation/thyroid nodules/feature fusion/deep learning/feature extraction/downsample/image pre-processing引用本文复制引用
出版年
2024
NETL
NSTL
万方数据