Lightweight Parasitic Egg Detection using Modified YOLOv11 with Ghost Convolutions, Hybrid SimAM-ECA Attention, and DropBlock Regularization

Abstract

Intestinal parasitic infections (IPIs) pose a significant global health challenge, particularly in resource-limited settings where traditional diagnostic methods are slow and labor-intensive. While deep learning models offer a promising solution for automated parasite egg detection, their large size often hinders deployment in real-time on low-power devices.. This study addresses the trade-off between accuracy and computational efficiency by proposing a lightweight and modified YOLOv11 architecture. We introduce a novel EnhancedConv block that integrates Ghost Convolution for efficiency, a hybrid attention mechanism combining SimAM and ECA-Net for improved feature focus, and DropBlock for robust regularization. We compared our Modified YOLOv11 Nano model (2.79 M parameters) against the baseline YOLOv11 Nano (2.62 M) and the larger YOLOv11 Small (9.45 M) on the public Chula-ParasiteEgg-11 dataset. Experimental results demonstrate that our modified model achieves a mAP@0.5 of 0.9627 and an F1-score of 0.9219, significantly outperforming the original Nano model and matching the performance of the Small model while requiring only ≈ 42 % of its computational cost (GFLOPs). This work demonstrates that strategic architectural enhancements can yield a model that is both highly accurate and efficient, presenting a viable solution for accessible, large-scale parasitic egg detection in resource-constrained environments.