基于SOANet的圆锥角膜辅助诊断

Keratoconus assisted diagnosis based on SOANet network

摘要:

亚临床期圆锥角膜(subclinical keratoconus,subkc)发病隐匿,现有医疗设备诊断存在局限性,所以提出一种诊断亚临床期圆锥角膜的检测方法是十分必要的。有研究发现圆锥角膜(Keratoconus,kc)力学性能的改变早于形态学,因此从角膜生物力学的角度筛查亚临床期的圆锥角膜更加符合临床实际。本文运用角膜生物力学特征,以点云数据作为网络输入数据,将SO-Net (self-organizing network)和自注意力(self-attention,SA)机制结合构建SOANet,对圆锥角膜、亚临床期圆锥角膜和正常角膜进行分类。首先,利用可视化生物力学分析仪(corneal visualization Scheimpflug technology,Corvis ST)拍摄角膜受力形变视频,对其进行处理得到点云数据集,接着对点云数据进行增强处理,使3种类型的角膜数据量分布均衡。然后按照3∶1的比例划分训练集和测试集, 分别对角膜进行二分类和三分类。最终模型在二分类和三分类测试集上的准确率分别达到98.3%和91.26%,即有效识别亚临床期圆锥角膜和圆锥角膜。实验结果表明,以三维点云数据构建亚临床期圆锥角膜辅助诊断模型是可行的,SOANet能够有效识别出亚临床期圆锥角膜,且分类性能优于传统模型。

Abstract:

The onset of subclinical keratoconus (subkc) is hidden,and existing medical equipment has limitations in diagnosis.Therefore,it is necessary to propose a detection method for diagnosing subclinical keratoconus.Studies have found that the mechanical properties of keratoconus (kc) change earlier than morphology,so screening subclinical keratoconus from the perspective of corneal biomechanics is more in line with clinical practice.This article utilizes corneal biomechanical features and uses point cloud data as network input data.Self organizing network (SO- Net) and self attention mechanism (SA) are combined to construct SOANet,which classifies keratoconus,subclinical keratoconus,and normal corneas.Firstly,a corneal visualization Scheimpflug technology (Corvis ST) was used to capture a corneal deformation video,which was processed to obtain a point cloud dataset.The point cloud data was then enhanced to achieve a balanced distribution of the three types of corneal data.Then the training set and test set were divided in a 3∶1 ratio,and the cornea was classified into two categories and three categories,respectively.The accuracy of the final model on the two categories and three categories test sets reached 98.3% and 91.26%,respectively,effectively identifying subclinical keratoconus and keratoconus.The experimental results indicate that constructing a subclinical keratoconus assisted diagnostic model using 3D point cloud data is feasible,and SOANet can effectively recognize subclinical keratoconus,and its classification performance is better than traditional models.