Enhancing Multi-Scale Representation in ResNest for Retinal Image Feature Extraction

The use of convolutional layers for retinal image feature extraction results in varying features extracted from different layers. Previous research has only utilized fixed-resolution images with single-layer features to train the network model, failing to fully utilize the diverse feature information available. To address this, we improved the multi-scale representation capability of the backbone by implementing hierarchical residual-like connections within each single radix block of the original ResNest. This involved replacing the feature maps convolution layer with smaller groups of 3x3 filters and connecting them hierarchically in a residual-like style across different filter groups within each radix of ResNest.