Fig. 7: Illustration of the architecture of Freeze-Missing module.

The Freeze-Missing module can model missing data for the pre-diagnosis of visual field defects, providing effective support in regions with limited medical resources. The numerical flow extracts high-dimensional features using fully connected layers. The feature extraction networks for RNFL flow, focused flow, and global flow share the same structure, consisting of convolution, max pooling, and Se-Conv layers. Se-Conv can determine the optimal image channels in an end-to-end manner. Image feature fusion is achieved with Se-Conv, and multimodal feature fusion is performed via concatenation. Finally, a fully connected layer is used for class mapping.