High-throughput image-based pupal sex classification in Aedes aegypti using convolutional neural network models for...

Aedes aegypti, the primary vector of dengue, yellow fever, Zika, and chikungunya, poses serious public health threats, especially in tropical regions. With limited vaccine availability, innovative control strategies like the Sterile Insect Technique (SIT) are essential, requiring efficient pupal sex separation to release only non-biting sterile males. This study developed a high-throughput, low-cost imaging pipeline using 96-well plates with optimized conditions (fixed camera stand, backlighting, refrigeration for immobilization, and black-painted wells) to capture mosquito pupal images using a mobile phone camera. Custom Python tools automated image preprocessing, segmentation, and labeling. Lightweight convolutional neural networks (CNNs), namely MobileNetV2, MobileNetV3-Large, and EfficientNetV2B0, were fine-tuned with transfer learning on datasets ranging from 2 to 290 image sets. MobileNetV3 achieved 94.17% accuracy at 290 sets, outperforming MobileNetV2 (91.77%) and EfficientNetV2B0 (88.96%), following power-law scaling with reduced variability in larger datasets. Confusion matrices revealed persistent false positives (females misclassified as males), critical for SIT to prevent disease transmission. Validation against human annotators (89.84% accuracy) and flatbug, a purely size-based automated segmentation model (81.70% accuracy), showed MobileNetV3 to be significantly superior (p<0.05). Furthermore, Gradient-weighted Class Activation Mapping confirmed that the CNN robustly learned complex morphological features rather than relying on size metrics. This workflow integrating fine-tuning strategies, dataset scaling, automated interpretability, and CNN-human validation addresses SIT bottlenecks in resource-limited settings and may extend to binary pupal sex classification of other mosquito species and strains for enhanced SIT application.