Dataset and Pre-trained Models for: A Bio-Consistent Hierarchical Cascade Network (BCHCN) for Bacterial Classification

This repository contains the essential dataset and pre-trained model weights required to reproduce the Bio-Consistent Hierarchical Cascade Network (BCHCN) proposed for automated bacterial classification. Contents: data.zip: Contains the hierarchical patch-level microscopy images utilized for training and evaluation. The data is structured to support multi-scale inference, specifically aligned for Gram staining, morphological feature extraction, and species-level identification tasks. models.zip: Contains the optimized ResNet-50 backbone weights for the three distinct hierarchical stages: Stage 1 (Gram constraint), Stage 2 (Morphological constraint), and Stage 3 (Species identification). models_convnext.zip: Contains the optimized ConvNeXt-Tiny weights utilized as a modern CNN baseline. The architecture was comprehensively trained and evaluated across multiple hierarchical crop levels (denoted Zoom 1-5, corresponding to different field-of-view crops), with checkpoints saved at peak validation accuracy. Since the Zoom 1 crop achieved the maximum overall validation accuracy, it was selected as the optimal input size for the final test-set evaluation, ensuring a methodology free of data leakage. This model serves as a comparative baseline to demonstrate the vulnerability of standard CNNs to morphologically impossible errors in the absence of hierarchical biological constraints. models_swin.zip: Contains the optimized Swin Transformer (Swin-T) weights utilized as a modern CNN baseline. The architecture was comprehensively trained and evaluated across multiple hierarchical crop levels (denoted Zoom 1-5, corresponding to different field-of-view crops), with checkpoints saved at peak validation accuracy. Since the Zoom 1 crop achieved the maximum overall validation accuracy, it was selected as the optimal input size for the final test-set evaluation, ensuring a methodology free of data leakage. This model serves as a comparative baseline to demonstrate the vulnerability of standard CNNs to morphologically impossible errors in the absence of hierarchical biological constraints. These resources are designed to be used in conjunction with the official GitHub repository associated with the manuscript, ensuring full reproducibility of the double-masking clinical safety evaluations and Grad-CAM feature attributions.