全局相关块级自注意力的食管癌前病变区域分割

Segmentation for Esophageal Precancerous Lesions Region Using Globally Correlated Block-Level Self-Attention

食管内镜图像多类癌前病变区域的类间特征差异小且个体差异大,难以实现高精度分割,使用自注意力机制可提取远距离依赖信息以获取判别性特征,但是计算开销大。为此,提出一种全局相关块级自注意力(GC-BLSA)方法,用于食管癌前病变区域分割。利用块级自注意力对骨干网络的输出特征进行分块,并在多个特征块上引入自注意力机制,降低网络参数量和计算量,采用块相关机制建模每个特征块和整个特征图之间的关系,解决每个特征块在单独使用自注意力时无法提取与全局相关的远距离依赖信息的问题。在此基础上,在块级自注意力模块中引入相对位置偏移对位置信息进行补充,有效提高网络分割精度。实验结果表明,在四分类食管癌数据集上该方法的分割指标mIoU和F1值分别为50.213%和63.786%,相比传统自注意力Non-local模块分别提高3.744和4.297个百分点,参数量和计算量分别下降26.38%和10.62%。

The inter-class feature differences and individual differences between multiple types of precancerous lesions in esophageal endoscopic images are small,making it difficult to achieve high-precision segmentation.The use of self-attention mechanism can extract long-distance dependent information to obtain discriminative features,but the computational cost is high.This study proposes the Globally Correlated Block-Level Self-Attention(GC-BLSA)method for the segmentation of esophageal precancerous lesions.Block-Level Self-Attention(BLSA)is used to block the output features of the backbone network,and self-attention mechanism is introduced on multiple feature blocks to reduce the amount of network parameters and computation.The Block Correlation Mechanism(BCM)is used to model the relationship between each feature block and the entire feature map,solving the problem of each feature block being unable to extract long-distance dependency information related to the entire local area when using self-attention alone.Based on this,a relative position offset is introduced in the BLSA module to supplement position information,effectively improving the network segmentation accuracy.The experimental results show that the segmentation indicators mIoU and F1 values of the proposed method on the four-classification esophageal cancer dataset are 50.213%and 63.786%,which are 3.744 and 4.297 percentage points higher than the traditional self-attention Non-local module,respectively.The parameter quantity and computation of the proposed method are reduced by 26.38%and 10.62%,respectively.