坐标注意力与可变形卷积增强的YOLO用于CT图像中双侧上颌窦（一种鼻旁窦结构）分割

文献信息

中文摘要

背景： 在CT图像中准确勾画上颌窦是支持耳鼻咽喉科术前规划和鼻窦疾病评估的相关步骤。由于周围解剖结构复杂、骨性结构菲薄且患者间差异大，手动勾画这些鼻旁窦结构困难且容易出错。本文提出了一种实例分割框架，通过整合坐标注意力（CoordAtt）模块和可变形卷积网络v2（DCNv2）来扩展YOLO11n-Seg架构，以提高CT图像中左右上颌窦的空间特征表示和边界精度。CoordAtt沿水平和垂直轴编码方向性空间上下文以增强解剖定位，而DCNv2将选定的固定卷积核卷积层替换为自适应可变形采样，以更好地捕捉不规则和薄壁的窦边界。方法：所提出的模型在公开可用的NasalSeg数据集上进行训练和评估，使用来自130例患者的包含任一侧上颌窦的所有轴位CT切片（3,620张切片；训练/验证/测试分别为104/13/13例患者），并与三个YOLO系列基线（YOLOv8n-Seg、YOLO11n-Seg和YOLO26n-Seg）进行比较。结果：在382张切片的测试集（13例患者）上，所提出的模型达到了93.07%的掩膜mAP50和74.49%的掩膜mAP50-95，分别比三个基线高出最多+4.60和+3.30个百分点，并且框mAP50为91.00%，mAP50-95为75.83%。组件消融研究证实，这一改进源于CoordAtt和DCNv2的联合使用，而非任一机制单独作用；患者级bootstrap置信区间和配对显著性检验证实了所观察到的增益具有统计稳健性。结论：这些结果代表了概念验证证据，表明坐标感知注意力和可变形卷积可以有效地整合到基于YOLO的实例分割流程中，用于CT双侧上颌窦勾画。

英文摘要

Accurate delineation of the maxillary sinuses in CT images is a relevant step supporting preoperative planning and the evaluation of sinonasal disease within otolaryngology. Manual contouring of these paranasal sinus structures is difficult and error-prone, since the surrounding anatomy is complex, bony structures are thin, and inter-patient variability is large. This paper proposes an instance segmentation framework that extends the YOLO11n-Seg architecture through the integration of Coordinate Attention (CoordAtt) modules and Deformable Convolutional Networks v2 (DCNv2), to improve spatial feature representation and boundary precision for the right and left maxillary sinuses in CT images. CoordAtt encodes directional spatial context along horizontal and vertical axes to enhance anatomical localisation, while DCNv2 replaces selected fixed-kernel convolution layers with adaptive deformable sampling to better capture irregular and thin-walled sinus boundaries. The proposed model was trained and evaluated on the publicly available NasalSeg dataset, using all axial CT slices containing either maxillary sinus from 130 patients (3,620 slices; 104/13/13 patients across training/validation/test), and compared against three YOLO-family baselines (YOLOv8n-Seg, YOLO11n-Seg, and YOLO26n-Seg). On the 382-slice test set (13 patients), the proposed model achieved a mask mAP50 of 93.07% and mask mAP50-95 of 74.49%, outperforming the three baselines by up to +4.60 and +3.30 percentage points, respectively, and a box mAP50 of 91.00% and mAP50-95 of 75.83%. A component-wise ablation study confirmed that this improvement arises from the joint use of CoordAtt and DCNv2 rather than from either mechanism alone, and patient-level bootstrap confidence intervals and paired significance testing confirmed the statistical robustness of the observed gains. These results represent proof-of-concept evidence that coordinate-aware attention and deformable convolution can be effectively integrated into a YOLO-based instance segmentation pipeline for bilateral maxillary sinus delineation from CT.