联合自适应核和Transformer的脊柱磁共振成像多类别分割网络

Multi category segmentation network for magnetic resonance imaging spine based on adaptive kernel and transformer

针对脊柱磁共振成像(magnetic resonance imaging,MRI)结构复杂,存在多余组织、噪声及伪像的问题,本研究设计了一种联合自适应核和Transformer的脊柱MRI实例多类别分割网络。以Swin Transformer作为骨干网络,通过引入稠密连接模块减少前向通道的信息丢失,以更好地捕获图像中的细节和局部信息。同时,为进一步捕获复杂空间的多尺度特征,采取自注意力核选择的方式构建跨尺度稠密连接,使模型在训练过程中能自适应学习到合适的卷积核尺寸,提高模型对不同尺度信息的感知能力,提高分割性能。通过在215例受试者的T2加权MRI图像2D切片上进行验证,实验结果显示,该算法的平均交并比(mean intersection over union,mIoU)、平均召回率(mean recall rate,mRecall)和平均骰子系数(mean dice coefficient,mDice)分别为82.63%、89.37%和88.85%。结果表明,本研究算法的分割性能较好,可实现脊柱MRI中椎体及椎间盘的精准分割,为临床医生提供辅助诊断工具。

Aiming at the problems of complex structure,redundant tissue,noise and artifacts in MRI of spine,an instance multi category segmentation network of spine MRI based on adaptive kernel and transformer was designed.The Swin Transformer was used as the backbone network.By introducing the dense connection module,the information loss of the forward channel was reduced,so as to better capture the details and local information in the image.At the same time,to further capture the multi-scale features of complex space,a self-attention kernel selection method was adopted to construct the cross-scale dense connection,so that the model could adaptively learn the appropriate convolution kernel size in the training process,then improved the perception ability of the model for different scale information and the segmentation performance.Experiments on 2D slices of T2 weighted MRI images of 215 subjects showed that the mean intersection over union(mIoU),mean recall rate(mRecall),and mean dice coefficient(mDice)reached 82.63%,89.37%,and 88.85%,respectively.The results show that this algorithm has excellent segmentation performance,and can realize the accurate segmentation of vertebral bodies and intervertebral discs in spinal MRI images,can provide an auxiliary diagnostic tool for clinicians.