The classification stage of the pipeline relies on three widely used convolutional neural network (CNN) architectures: DenseNet121, ResNet50 and VGG16.

DenseNet121 promotes feature reuse via dense connectivity, which helps capture fine-grained parasite morphology. ResNet employs residual connections to train deeper models stably, and VGG16 acts as a strong classical baseline.

In the front-end simulation, these backbones are not re-trained but are represented by configurable options that control how the downstream federated optimisation and uncertainty-aware aggregation are visualised.