An attention framework based on multiple color spaces for artifact classification in whole slide images

Computational pathology models are sensitive to artifacts arising from tissue preparation and
whole-slide imaging. Common artifacts-including tissue folds, uneven staining, ink markings,
blur, and contamination-alter both structural and chromatic characteristics of histopathological
images, adversely affecting model reliability. To address this, we present AC-AttnNet, a multicolor-space network for robust artifact classification. Input images are transformed into RGB
(Red-Green-Blue), HSV (Hue-Saturation-Value), HED (via Hematoxylin-Eosin deconvolution),
and CIELAB (L*a*b*) color spaces, from which features are extracted using Attention-Enhanced
Residual Blocks (AERBs). A Cross-Color Attention Module (CCAM) models inter-space feature
dependencies through sample-adaptive 4 × 4 attention matrices, enabling dynamic fusion of colordomain representations. AC-AttnNet achieves a multi-seed mean test accuracy of 97.91 ± 0.16%
(best single run: 97.96%) and 99.10% validation accuracy on a breast histopathology artifact
dataset with only 7.7M parameters and 1.58G FLOPs. Cross-dataset evaluation demonstrates
strong generalization with an average accuracy of 87.3%, outperforming AR-Classifier (81.1%) and
ResNet-50 (76.7%). Ablation studies confirm that multi-color-space modeling improves accuracy
by +0.72% over the RGB-only baseline, and CCAM fusion yields modest but consistent gains
over static concatenation across seeds (+0.08% mean improvement), with stronger qualitative
support from t-SNE silhouette improvement and Grad-CAM focus. These findings demonstrate
the effectiveness of cross-color feature interaction for artifact classification and support the use of
AC-AttnNet as a reliable component in computational pathology quality control and restoration
pipelines.