圆锥角膜是一种特征为角膜向前凸起呈圆锥形的眼病，常导致视力下降并影响日常生活，甚至需要进行角膜移植手术。因此，其早期诊断十分重要。在我国临床工作中，医生常使用角膜地形图等检查辨别早期圆锥角膜，但准确诊断潜在的圆锥角膜仍是挑战。因此，借助计算机深度学习技术辅助筛查和诊断具有现实意义。

本文以圆锥角膜的识别为例，研究了基于深度学习的图像识别在医疗影像中的应用。本文使用了残差神经网络ResNet50的结构建立分类模型，采用两阶段学习的方法，将经过预处理的角膜地形图分为圆锥角膜、正常角膜和疑似圆锥角膜三类。在训练过程中，本研究根据样本的特点，添加了数据增强、随机失活、可变学习率等方法，选用了焦点损失函数降低类别不平衡的影响，并用随机梯度下降法优化模型参数。最终，模型分类的准确率为98%，圆锥角膜、正常角膜和疑似圆锥角膜三个类别的精确率分别为98.04%，100%和96.08%，灵敏度分别为100%，96%和98%。模型的整体分类效果较好，对圆锥角膜的大量筛查等临床场景可以起到一定的辅助参考作用。

Keratoconus is an eye disease characterized by a cone-shaped cornea that bulges outward. It often causes blurred vision and has negative impacts on daily life, even requiring a corneal transplantation. Thus, it is important to have a correct diagnosis in the early stage. In clinical practice in China, corneal topography is often used to detect keratoconus, but it is still a challenge to diagnose potential keratoconus accurately. Therefore, it is of great significance to apply computer and deep learning technology to assist screening and diagnosis.

In this paper, by taking keratoconus as an example, the application of image recognition using deep learning in medicine was investigated. An image recognition model for keratoconus was established, classifying the images of corneal topography into three categories: keratoconus, normal cornea, and suspected keratoconus. This model used the structure of ResNet50, a kind of convolutional neural network, and included two learning stages of binary classification process. During the training process in this study, Focal Loss and Stochastic Gradient Descent were used to optimize the parameters of the model, combining methods like data augmentation, dropout, and variable learning rate. After evaluation, the accuracy of this model was 98%, the precision of three categories of keratoconus, normal cornea, and suspected keratoconus was 98.04%, 100%, and 96.08% respectively, and the sensitivity was 100%, 96%, and 98% respectively. The results showed that this model with good performance could assist ophthalmologists in screening keratoconus.