Classification of Foot-and-Mouth and Lumpy Skin Disease in Cattle Using Frozen Pretrained Encoders: A Comparison of Convolutional and Transformer Architectures

Objective: To determine whether the architectural family of a pretrained encoder, convolutional or attentional, governs its transferability and the reliability of its probabilities under frozen transfer, and how much of the task can be solved once lesion morphology is removed from the input. Methods: Photographs of cattle labelled healthy, foot-and-mouth disease or lumpy skin disease were split into a development set of 1277 from a public Kaggle collection and an external set of 985 from two Zenodo records. Six encoders, three convolutional and three vision transformers (ViT-B/16), were frozen and read through an identical linear head under a grouped five-fold split with five seeds. Discrimination and calibration were co-primary, and the external set was read once. Results: External balanced accuracy ran from 0.906 to 0.819, both extremes being ViT-B/16 encoders; after Holm correction, the three leading encoders, two attentional and one convolutional, were indistinguishable. The lowest of the six returned a Brier skill score of −0.030. A 4 × 4 thumbnail reached 0.527 against 0.333 for chance and recovered 0.650 of the foot-and-mouth cases. Conclusions: The pretraining recipe, not the convolutional–attentional distinction, ordered both criteria. Discrimination gave no warning of the loss of probabilistic skill, so a proper scoring rule must be reported beside it. External validation cannot expose a cue present in both collections that travels with the label, and low-information controls must accompany it.