边缘增强的变感受野自注意力撕囊评估算法

为了解决前囊的透光性、器械遮挡和较大的场景差异造成的分割困难，实现撕囊操作的定量评估、缩短医生的学习曲线、提供更规范的治疗，通过边缘增强的方法使得结合了自注意力的变感受野网络关注边缘特征，更加准确地分割出撕囊形成的前囊孔洞并对其进行定量评估。首先利用变感受野的空洞卷积网络获取多尺度的语义依赖，接着通过自注意力机制建立边缘和目标内部的依赖关系，最后利用分割结果计算出用于手术评估的圆度、居中度和半径指标。自建了白内障撕囊数据集，并利用球面化算法对其进行了增广和误操作模拟，提出的分割算法在该数据集上进行了测试，实验结果表明所提出的算法明显优于传统分割算法，精确度、交并比分别达到了96.51%、93.25%。可见，该算法能够实现更精准的前囊孔洞分割和术后评估。

Boundary Enhanced Self-Attention with Variable Receptive Field for Capsulorhexis Assessment

In order to address the segmentation challenges caused by the translucency of the anterior capsule, device occlusion and large scene differences, to achieve quantitative assessment of the capsulorhexis, shorten the learning curve of the surgeon and provide a more standardised treatment, a boundary enhancement pipeline was used to enable a variable perceptual field network incorporating self-attentiveness to focus on boundary features and to more accurately segment for quantitatively assessment. A multi-scale semantic dependency is first obtained using a dilation convolutional network with variable receptive field, followed by a self-attention mechanism to establish the dependency between the boundary and the interior, and finally the segmentation results are used to calculate the roundness, centrality and radius metrics for surgical evaluation. The proposed segmentation algorithm was experimented on a self-built capsulorhexis dataset and spherification algorithm was used for augmentation and misoperation simulations. Experimental results show that the proposed algorithm significantly outperforms traditional segmentation algorithms by 96.51%Dice and 93.25%IoU. It is evident that the algorithm is capable of achieving more accurate anterior capsule hole segmentation and postoperative assessment.