WaveCAF: A Wavelet–Spatial Cross-Attention Fusion Network for Robust and Data-Efficient Pneumonia Detection in Chest X-Rays

Automated pneumonia screening from chest X-rays (CXRs) faces two coupled difficulties. Annotated radiographs are scarce. The disease also presents in two forms at once: focal lobar consolidation and diffuse interstitial infiltration. Purely spatial encoders resolve compact, high-contrast consolidations well. They under-exploit, however, the multi-scale frequency texture that carries most of the discriminative energy of diffuse infiltrates. We introduce WaveCAF, a dual-branch network that couples a ConvNeXt-Tiny spatial encoder with a frequency branch built on a single-level 2D discrete wavelet transform (DWT). The four Haar sub-bands (LL/LH/HL/HH) are re-weighted by an energy-driven learnable sub-band attention, encoded into tokens, and merged with the spatial tokens through a bidirectional Cross-Attention Fusion (CAF) block whose outputs are combined by a learned scalar gate; training couples a class-balanced focal loss with a cross-branch spectral-consistency regularizer, so class imbalance is handled analytically rather than by synthetic oversampling. On the Kaggle Chest X-ray Pneumonia benchmark, the released reference implementation of WaveCAF attains 97.9% (±0.4) accuracy, 98.5% F1, 98.3% recall and an AUC of 0.998 under stratified 5-fold cross-validation, statistically indistinguishable from the ConvNeXt-Tiny spatial baseline trained under the identical protocol (98.1% ± 0.2) and at the level that our harmonized reproductions identify as the ceiling of this benchmark. We further report (i) an ablation that isolates the sub-band attention, the cross-attention fusion and the gate against element-wise-summation and channel-concatenation fusion controls, (ii) a two-model data-efficiency sweep from 10% to 100% of the labels, (iii) a computational-efficiency budget covering parameters, FLOPs and single-image GPU latency, (iv) a robustness evaluation under thirteen image-degradation settings, and (v) an external evaluation on 26,684 adult films of the multi-institutional RSNA cohort, in which both models fall to ~0.87 AUC and the fused model detects more pneumonias at equal AUC (recall 69.6% vs. 62.7%, balanced accuracy 77.0% vs. 75.4%), together with a qualitative Grad-CAM and sub-band saliency analysis. On clean, fully supervised data the fused model does not surpass the spatial baseline; all ablated configurations remain within the noise floor of this saturated task. Once its gate and frequency encoder are properly initialized, however, WaveCAF is consistently at or above the baseline at every labeled-data fraction. It is also markedly more robust to degradations that suppress spatial amplitude: up to +7.8 accuracy points under severe contrast reduction, +1.9 under blur, and a higher AUC in all thirteen corruption settings. The cost is an increased sensitivity to additive noise, which contaminates the detail sub-bands directly. We report both directions of this trade-off in full, and we delimit where an explicit frequency prior pays.