<![CDATA[This study evaluates the integration of attention mechanisms (Efficient Channel Attention [ECA] and Coordinate Attention [CoordAtt]) into MobileNetV3-Small for preliminary conjunctival anemia screening. Using a single public dataset of 710 pediatric images from Ghana (CP-AnemiC dataset), the model was evaluated retrospectively via stratified 10-fold cross-validation (80%) and tested on a hold-out test set (20%). CoordAtt achieved the highest performance (Accuracy 78%, Sensitivity 84%, Specificity 70%, AUC 0.85), outperforming the Baseline (Accuracy 75%, Sensitivity 80%, Specificity 67%, AUC 0.80) and ECA (Accuracy 76%, Sensitivity 83%, Specificity 66%, AUC 0.81). A 95% Bootstrap BCa confidence interval test confirmed that the CoordAtt improvement over the Baseline is statistically significant for Accuracy, F1-Macro, Sensitivity, and AUC. Grad-CAM visualizations further showed more focused conjunctival localization. This improvement came with only a minimal increase in theoretical computational footprint (1.06M parameters, 0.0584 GFLOPs), keeping a lightweight profile similar to the Baseline (1.03M parameters, 0.0581 GFLOPs). These results show that CoordAtt improves classification performance without sacrificing computational efficiency, suggesting potential suitability for resource-constrained devices prior to clinical validation.]]>
Copyrights © 2026
