Original-image-controlled evaluation of deep and hHybrid models for bacterial morphology classification in gram-stained microscopy

Abstract

Bacterial morphology assessment from Gram-stained microscopy is a relevant visual task in clinical microbiology, especially in ROI-based datasets where multiple regions may originate from the same image. This study focuses on leakage-controlled morphology classification, treating Gram status as contextual information because its performance was near ceiling. Using 1,958 microscopy images with ROI-level annotations, we compared an original-image-controlled partition, which kept all ROIs from the same image within a single subset, with an ROI-level split that allowed image overlap across subsets. We evaluated explicit morphological descriptors, EfficientNet-B0, MobileNet, and hybrid probability- and embedding-level fusion strategies. Gram classification reached F1 values close to 1.0000, limiting its comparative value. In contrast, ROI-level splitting increased the Macro-F1 by up to 0.0375. Deep and hybrid models substantially outperformed the morphology-only baseline, but adjusted statistical tests did not support robust hybrid superiority over the best deep-only model.