基于TransUNet的甲状腺结节超声图像精准分割方法  

TransUNet-based method for accurate segmentation of ultrasound images of thyroid nodules

作　　者：陈格 李翔[1] CHEN Ge;LI Xiang(Jingbei Medical District,Chinese PLA General Hospital,Beijing 100080)

机构地区：[1]解放军总医院京北医疗区,北京100080

出　　处：《北京生物医学工程》2024年第2期165-170,共6页Beijing Biomedical Engineering

摘　　要：目的甲状腺结节的精准分割在医学影像处理中具有重要意义,然而,超声图像中的结节通常具有尺寸多变和边缘模糊的特点,这为其准确分割带来了挑战。为有效应对这一挑战,本文提出了一种结合卷积神经网络(convolutional neural network,CNN)和Transformer的分割网络,命名为TransUNet,旨在实现对甲状腺结节超声图像的精准分割。方法首先,使用卷积神经网络对超声图像进行编码,以生成特征图。接着,将特征图转换为序列向量,并利用Transformer的编码机制来捕捉上下文信息。此外,为保持局部细节特征的完整性,研究组还引入了跳跃连接,将其用于在解码器中对编码特征进行上采样,这对于处理边缘模糊等问题尤为重要。结果通过在甲状腺结节图像分割任务中进行广泛的实验,验证TransUNet的有效性。具体而言,骰子系数(dice coefficient,DICE)为0.75,交并比(intersection over union,IoU)为0.60,F1分数(F1 Score)为0.72,准确率高达0.93,AUC(area under the ROC curve)为0.91。这些性能指标反映了该方法在处理尺寸多变和边缘模糊等挑战方面的出色表现。结论本文提出的TransUNet为甲状腺结节超声图像分割任务带来了显著的性能提升。相较于传统的U-Net方法,TransUNet不仅更好地处理了尺寸多变和边缘模糊等挑战,而且在分割性能上具有更为出色的表现,为医学图像处理领域的进一步研究和临床应用提供了有力支持。Objective The precise segmentation of thyroid nodules holds significant importance in medical image processing.However,nodules within ultrasound images often exhibit variable sizes and blurred edges,posing a challenge to their accurate segmentation.To effectively address this challenge,this study introduces a segmentation network that combines convolutional neural networks(CNN)and Transformer,named TransUNet,with the goal of achieving accurate segmentation of thyroid nodule ultrasound images.Methods First,the ultrasound images are encoded by using CNN to generate feature maps.Subsequently,these feature maps are transformed into sequence vectors,and the encoding mechanism of the Transformer is employed to capture contextual information.Additionally,to preserve the integrity of local detail features,we introduce skip connections,which are utilized in the decoder to upsample the encoded features.This step is particularly crucial for addressing issues such as edge blurriness.Results The effectiveness of TransUNet is confirmed through extensive experiments in the context of thyroid nodule image segmentation tasks.Specifically,the dice coefficient(DICE)achieves an impressive score of 0.75,the intersection over union(IoU)reaches 0.60,the F1 Score is as high as 0.72,and the accuracy reaches an outstanding 0.93.Furthermore,the area under the ROC curve(AUC)stands at 0.91.These performance metrics underscore the excellent performance of the proposed method in addressing challenges related to variable nodule sizes and edge blurriness.Conclusions The proposed TransUNet in this study has yielded significant performance improvements in the task of thyroid nodule ultrasound image segmentation.Compared to the traditional U-Net method,TransUNet not only excels in addressing challenges related to variable nodule sizes and edge blurriness but also exhibits superior performance in terms of segmentation accuracy.This advancement provides robust support for further research and clinical applications in the field of medical image proces

关 键 词：甲状腺结节 超声图像分割 深度学习 全局自注意力 跳跃连接 

分 类 号：R318.04[医药卫生—生物医学工程]