Comparative analysis of resource-efficient YOLO models for rapid and accurate recognition of intestinal parasitic eggs in stool microscopy

Abstract

Faster and reliable recognition of the specific species of intestinal parasite eggs in stool microscopic images is required for targeted and quick intervention of soil transmitted helminths (STH) disease. The main objective of the proposed work is to identify the effective light weight basic yolo models among the recent compact yolo variants such as yolov5n, yolov5s, yolov7, yolov7-tiny, yolov8n, yolov8s, yolov10n and yolov10s, that could assist in rapid and accurate recognition of 11 parasite species egg. The real time performance of the compact yolo models have been analyzed in embedded platforms: Raspberry Pi 4, Intel upSquared with the Neural Compute Stick 2 and Jetson Nano. Finally, Gradient-weighted class activation mapping (Grad-CAM) has been used as an explainable AI (XAI) visualization method to elucidate the egg detection performance of the proposed models. Yolov7-tiny achieved the overall highest mean Average Precision (mAP) score of 98.7 %. On contrary, yolov10n yielded highest recall and F1 score of 100 % and 98.6 %. On other hand, yolov8n took least inference time with processing speed of 55 frames per second with Jetson Nano. Notably, the proposed framework demonstrates superior performance in detection of egg classes - Enterobius vermicularis, Hookworm egg, Opisthorchis viverrine, Trichuris trichiura, and Taenia spp. which is a significant outcome of the current research. Further, Grad-CAM depicts the discriminative power of unique features in parasite eggs. Thus, this study demonstrates the effectiveness, compactness and inference latency analysis of basic compact yolo variants in learning the specific patterns, texture and shape of parasitic egg species, thereby potentially enhancing the diagnostic accuracy of STH.