Two-Stage Hierarchical Pruning (THP-CNN) of Convolutional Neural Networks for Rapid Pathogenic Bacterial Detection Using High-Resolution Colony Images in Intensive Care Units

Background/Objectives: Patients in Intensive Care Units (ICUs) have an elevated risk of infection. Accurate identification of pathogenic bacteria is critical for targeted interventions; however, convolutional neural networks (CNNs) face challenges of high computational demands and parameter redundancy. Methods: We developed a two-stage hierarchical pruning framework for CNN compression (THP-CNN), combining channel importance estimation with receptive field equivalence transformation for a 24-class pathogenic bacteria classification task. Results: THP-CNN (70% pruned) achieves an accuracy of 0.86 with 0.62 M parameters, outperforming ResNet-50 (0.72), MobileNet V2 (0.81), Inception (0.74), and AlexNet (0.62), with the 50% and 60% pruned variants in cross-validation stably maintaining a mean accuracy of 0.79. Conclusions: THP-CNN demonstrates potential for lightweight, real-time bacterial classification, offering a computationally efficient solution for automated pathogen detection.