Ocular axial length (AL) is an important property of eyes used for determining their health prior to surgery. Estimation of AL is also crucial while making artificial lenses to replace impaired natural lenses. However, accurate measurement of AL requires a costly and bulky benchtop optical system. The complex structural features of eyes can be captured by fundus images, which can be easily captured nowadays with portable cameras. Here, we suggest a deep learning method for predicting AL based on fundus images with evidence of decision. This visual interpretation of predictions is achieved by post-processing, separated from the training process, to ensure that the architecture can be freely designed. Through the visualization technique, discriminative regions on input images can be localized to demonstrate specific areas of interest for predictions. In the experiments, we found a significant relationship between the fundus images and AL with achieving a coefficient of determination (R2) of 0.67 and accuracy of 90%, within an error margin of $ \pm 1$ mm. Furthermore, visual evidence proves that the network uses consistent regions for predicting AL. The visual results of this study also point to a link between AL and biological structure of eyes, which paves the way for future research.

中文翻译：

---

基于深度神经网络视觉解读视网膜眼底图像的眼轴长度预测

眼轴长度 (AL) 是眼睛的重要属性，用于在手术前确定其健康状况。在制造人工晶状体以替代受损的自然晶状体时，AL 的估计也很重要。然而，精确测量 AL 需要昂贵且笨重的台式光学系统。眼底图像可以捕捉眼睛复杂的结构特征，现在便携式相机可以很容易地捕捉到这些图像。在这里，我们提出了一种基于具有决策证据的眼底图像预测 AL 的深度学习方法。这种对预测的视觉解释是通过后处理实现的，与训练过程分开，以确保架构可以自由设计。通过可视化技术，可以对输入图像上的判别区域进行定位，以展示预测的特定感兴趣区域。在实验中，我们发现眼底图像和 AL 之间存在显着关系，确定系数 (R2) 为 0.67，准确度为 90%，误差范围为 $\pm 1$ mm。此外，视觉证据证明网络使用一致的区域来预测 AL。这项研究的视觉结果还指出了 AL 与眼睛生物结构之间的联系，这为未来的研究铺平了道路。