一种面向口腔移植骨分割的SA-UNet网络

SA-UNet neural network for oral grafted bone segmentation

在牙齿种植治疗中,口腔移植骨分割对于辅助医生诊断有重要参考价值。口腔CBCT图像具有对比度较低,移植骨边缘模糊等诸多特征,严重制约着现有深度分割网络的应用。研究以U-Net为基准网络,通过设计一种新颖的轻量级Sharp-Attention模块,提出了一种改进的SA-UNet网络模型。具体地,在Sharp模块中,带锐化卷积核的Depthwise卷积操作通过锐化浅层特征以加强特征细节。CBAM模块提升模型对于图像低层信息的关注度。进而,采用新型联合损失函数,缓解样本比例失衡带来的影响。最后,在口腔移植骨数据集上验证了模型的有效性。在模型复杂度方面,与基准网络U-Net相比,在几乎没有增加计算开销的情况下,图像分割精度得到了有效的提升;在分割精度方面,与现有的主流医学分割模型对比,在IoU、Dice系数、Hausdorff距离三个评价指标上的表现最佳,得分达到了0.8665、0.9262、0.5092。

In dental implant treatment,oral grafted bone segmentation has important value for assisting doctors in diagnosis.Oral CBCT images have many features such as low contrast and blurred edges of bone and restrict the application of the segmentation networks.Taking U-Net as the benchmark network,an improved SA-UNet model is proposed by designing a novel lightweight Sharp-Attention module.Specifically,in the Sharp module,the depthwise convolution with a sharpening convolution kernel enhances feature details by sharpening shallow features.The CBAM improves the model's attention to the low-level information.Furthermore,a novel joint loss function is adopted to alleviate the imbalance of the samples.Finally,the effectiveness of the model is validated on the dataset.In terms of complexity,compared with U-Net,the segmentation accuracy has been effectively improved with almost no increase in computational overhead;compared with state-of-the-arts,the best performance is achieved on IoU,Dice coefficient and Hausdorff distance,with scores reaching 0.8665,0.9262,and 0.5092.