A hybrid convolutional neural network–transformer architecture with cross-attention fusion for honeybee disease classification

Image-based classifiers proposed for honeybee brood disease monitoring are rarely examined for the possibility that they separate classes by acquisition artefacts rather than by lesion appearance. This study evaluates a compact hybrid Convolutional Neural Network (CNN)–Transformer classifier, LGCA-Net (Local-query Global-key Cross-Attention Network; 1,459,074 parameters, trained from scratch), on two binary tasks, Chalkbrood and Foulbrood, each built from 1000 images of the public AI-Hub Honeybee Disease Diagnosis Image Dataset, and it audits the data before any model is interpreted. The audit showed that global image statistics differ strongly between classes (CIELAB L ∗ , Hedges’ g=−2.48 for Chalkbrood), that disease-positive images come from one and three proxy acquisition sites against thirteen for normal images, and that classifiers using only six global statistics reach balanced accuracies of 0.909–0.947 (Chalkbrood) and 0.841–0.900 (Foulbrood), inside the band later occupied by the deep networks. All comparisons were therefore repeated on acquisition-matched subsets (same sites, paired on lightness) under stratified hive-grouped three-fold cross-validation with a per-model learning-rate search and fold-wise DeLong and exact McNemar tests with Holm correction, against nine ImageNet-pretrained baselines (ResNet-50, DenseNet-121, EfficientNetV2-S, CoAtNet-0, CrossViT-S, Next-ViT-S, Conformer-S, LSNet-T, MambaVision-T). Numerically, on the original subsets LGCA-Net reached a balanced accuracy of 0.951 (95% CI 0.938–0.964) for Chalkbrood, within 0.016 of the best baseline (MambaVision-T, 0.967), and 0.894 (0.875–0.912) for Foulbrood, 0.080 below the best baseline (Conformer-S, 0.974). Matching reduced the standardised mean difference in L ∗ from −2.48 to −0.42 and from −1.35 to −0.04 and removed the single-threshold lightness cue (0.909 to 0.592; 0.841 to 0.417). Every deep network then lost accuracy and the ordering changed: LGCA-Net fell to 0.798 (0.711–0.884) and 0.773 (0.657–0.907), below all nine pretrained baselines in both tasks and 0.087 and 0.172 below MambaVision-T (0.885) and Conformer-S (0.945); on the matched subsets the comparisons with the pretrained baselines reached significance after correction only in the largest fold of each task and always favoured the baseline. An ablation supports the two-path design but not the cross-attention module, which was not distinguishable from position-wise additive fusion. A per-model analysis then characterises the association behind the ordering change: after matching, the decisions of LGCA-Net remain more reproducible from six global acquisition statistics than the disease label itself is (＋0.114 and ＋0.176 balanced-accuracy points above that reference), whereas the three plain pretrained CNNs and MambaVision-T stay within −0.025 to ＋0.044 of it. Because the matched subsets remain partially imbalanced in colour and file size and the learning rate was selected on a single validation split, the matched-subset comparisons are reported as exploratory. The contributions are methodological: a documented audit showing that the naive public subsets are confounded, a matched-subset protocol that changes the model ordering, an annotation-free measure of each network’s association with the acquisition channel, and a parameter-compact architecture, released with all code, derived data and per-image predictions, that does not exceed the larger pretrained networks. • A 1.46 M-parameter CNN–Transformer with cross-attention fusion for brood disease. • Brightness statistics alone reach 0.91–0.95 balanced accuracy on the public subsets. • Site-restricted, lightness-matched subsets remove the cue and change model ordering. • Under matching the compact model is below all nine larger pretrained baselines. • Ablation supports the two-path design but not the cross-attention fusion module. • Six scalars reproduce each network’s own decisions, quantifying confound reliance.