Microscopic detection of nematodes in entomopathogenic nematode-enriched samples using a lightweight deep learning model.

Automated detection of entomopathogenic nematodes (EPNs) is increasingly important in biological control research, where manual microscopic counting remains labour-intensive and prone to variability. Building upon existing computer vision approaches but differing in architectural design and computational efficiency, this study presents LightDetectorMS, an ultra-lightweight, anchor-free object detection framework optimized for microscopic imagery of laboratory-isolated Steinernema feltiae infective juveniles. The model was evaluated using five-fold cross-validation to assess reliability and generalizability. LightDetectorMS achieved a mean mAP@0.5 of 0.9119 (±0.0242) and a mean mAP@0.5:0.95 of 0.8207 (±0.0353), with precision and recall of 0.9184 (±0.0227) and 0.9382 (±0.0452), respectively, demonstrating stable performance across folds. The coefficient of variation remained below 5% for all metrics, supporting statistical consistency. The architecture contains only 62,991 parameters (0.46 MB) and operates at 152.5 FPS (6.56 ms per frame), enabling real-time processing even in dense microscopic fields with overlapping individuals. Comparative analysis with manual expert counting (n = 50) revealed a mean human counting time of 43.08 ± 3.00 s per sample, corresponding to 0.483 ± 0.03 nematodes per second. In contrast, LightDetectorMS processes equivalent workloads several thousand times faster while maintaining high detection reliability. These findings indicate that LightDetectorMS provides a computationally efficient and statistically robust solution for semi-automated quantification of EPNs in controlled laboratory environments, supporting large-scale biological monitoring and production workflows.