Deep learning has demonstrated high efficiency in histopathological image analysis, particularly in lung cancer classification. However, the stability of these models with image corruption and cross-dataset validation remains an important practical concern. In this study, we explored the potential of adding spectral information derived from the discrete wavelet transform (DWT) and spatial convolutional representations to enhance the robustness of multi-class lung cancer classification between Normal, Adenocarcinoma and Squamous cell carcinoma. The lightweight ResNet18 backbone was used to obtain spatial features, and spectral descriptors were obtained through wavelet sub-bands and integrated through early feature-level fusion. The models were trained and evaluated using the LC25000 dataset. Subsequently, it was tested under controlled perturbations, such as Gaussian noise and Gaussian blur. Three random seeds were used to assess performance variability, and paired t-tests were conducted as an indicative statistical measure of the results. Under clean conditions, the spatial and hybrid models were nearly saturated, and there was no significant difference between them (spatial: 99.85 ± 0.26; hybrid: 99.72 ± 0.22; p = 0.1217). The hybrid model exhibited higher robustness when Gaussian noise (σ = 0.05) was added, which resulted in 84.89% ± 4.52% accuracy versus 74.99% ± 7.20% of the spatial baseline (p = 0.0443) with an observed effect size (Cohen’s d = 2.64), noting that these estimates are based on a limited number of runs and should be interpreted with caution. The same behavior was observed in Gaussian blur perturbations, where the hybrid representation was slightly more stable. We also investigated a simplified adaptive gating mechanism process and found that the learned gate parameter also tends to converge towards spatial feature dominance with a model trained with clean data. Finally, cross-dataset validation with LungHist700 showed a slight increase in the balanced accuracy of the hybrid model (0.5158) over the spatial baseline (0.4722). These results indicate that spectral and spatial features can be used to enhance robustness to image corruption and still yield high classification accuracy, indicating that spectral–spatial representations can improve robustness under controlled perturbations, whereas their impact on cross-dataset generalization remains limited. The results further indicate that robustness improvements are strongly influenced by training strategies, such as noise augmentation, whereas the contribution of fusion is comparatively moderate.
Illa et al. (Fri,) studied this question.