🧬 biology

Bacterial Species Classification from Full-Plate Culture Images Using Deep Learning: A Comparative Study of CNNs and Vision Transformers

This study demonstrates that fine-tuned Vision Transformers (ViT-B/16) outperform traditional CNN architectures like ResNet50 and DenseNet121 in accurately classifying 19 bacterial species directly from full-plate culture images, achieving a mean accuracy of 99.05% under a unified experimental protocol.

Original authors: Abdulsalam Abdulsattar Abdulazez, Hussein Salim Abed, Aythem Khairi Kareem, Mustafa Nadhim OWAID, Ahmed Adil Nafea

Published 2026-10-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abdulsalam Abdulsattar Abdulazez, Hussein Salim Abed, Aythem Khairi Kareem, Mustafa Nadhim OWAID, Ahmed Adil Nafea

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the quiet hum of a clinical laboratory, a petri dish sits under a soft light, its surface dotted with tiny, growing colonies of bacteria. For decades, identifying what these invisible invaders are has relied on the trained eye of a microbiologist. Experts look at the size, color, texture, and shape of these colonies, comparing them against years of experience to name the species. This process is vital; knowing exactly which bacterium is causing an infection allows doctors to choose the right antibiotic quickly, saving lives and preventing the spread of disease. Yet, this human expertise is a bottleneck. It takes time, it is tiring, and different experts might even disagree on what they see. The question facing modern science is whether a computer can learn to see these patterns just as well, or perhaps even better, by looking at the entire dish at once rather than focusing on a single spot.

A team of researchers has now taken a significant step toward answering that question. They set out to build a system that could automatically identify nineteen different types of clinically important bacteria directly from photographs of full culture plates. Instead of asking a computer to look at a single, isolated colony, the team fed it images of the entire dish, capturing the complex arrangement of all the bacteria growing together. They tested three different types of artificial intelligence models to see which one could learn the visual language of these microbes most effectively. Two of the models were based on a design that mimics how the human eye processes images in layers, scanning for edges and shapes. The third model used a newer approach that looks at the whole picture at once, weighing how different parts of the image relate to one another.

The researchers worked with a carefully prepared collection of 950 high-resolution photographs. Each photo showed a petri dish containing one of nineteen specific bacterial species, such as Escherichia coli or Staphylococcus aureus, all grown under strict, identical conditions to ensure the images were fair and consistent. To train the computers, the team showed them these images over and over, making slight, random changes to the pictures—flipping them, rotating them, or adjusting the colors—to teach the models to recognize the bacteria regardless of how they were positioned or lit. This process ensured the system learned the true identity of the bacteria, not just the specific way a photo was taken.

When the team put the three models to the test, the results were striking. The model that looked at the whole image at once, using a method called a Vision Transformer, proved to be the most accurate. It correctly identified the bacterial species in nearly 99 percent of the cases. The other two models, which relied on the layered, edge-detecting approach, also performed well, with one achieving about 97.6 percent accuracy and the other around 94 percent. The difference might seem small in percentage points, but in a medical setting, it represents a significant gap in reliability. The top-performing model made only nine mistakes across the entire set of 950 images, whereas the others made many more.

What is perhaps most interesting is not just that the new model won, but how it handled the difficult cases. Some bacteria look so similar to the human eye that even experts can struggle to tell them apart. The researchers found that the top model was better at distinguishing between these tricky pairs than the other systems. For instance, it rarely confused two specific types of bacteria that the other models often mixed up. This suggests that by looking at the entire plate and understanding how the different colonies are arranged and interact, the model can pick up on subtle clues that a simpler, piece-by-piece analysis might miss.

The study does not claim to have solved every problem in bacterial identification. The researchers were careful to note that their system was trained on images taken in a single, controlled laboratory setting. They acknowledge that real-world hospitals use different cameras, lighting, and growing conditions, which could confuse the system. Furthermore, the current test did not fully account for whether the system could handle completely new strains of bacteria it had never seen before. However, the work provides a clear and reproducible benchmark. It demonstrates that deep learning, particularly the newer transformer-based models, can achieve a level of precision that rivals human experts when looking at full culture plates.

This research offers a glimpse into a future where the initial screening of infections could be faster and more consistent. By proving that a computer can learn to read the visual story of a full petri dish with near-perfect accuracy, the study opens the door for tools that could assist doctors and lab technicians. While the technology is not yet ready to replace human judgment in every hospital, it shows that the path forward lies in systems that understand the whole picture, not just the parts. The ability to identify these microscopic threats with such speed and accuracy could eventually mean that patients receive the correct treatment sooner, turning a slow, manual process into a swift, automated one.

Technical Summary: Bacterial Species Classification from Full-Plate Culture Images Using Deep Learning

Problem Statement

The morphological classification of microbial cultures on agar plates traditionally relies on specialized expertise and practical experience, a process that is time-consuming, labor-intensive, and prone to inconsistent interpretations among clinical technologists. While rapid and accurate identification of pathogens is critical for early infection diagnosis and antimicrobial treatment, existing automated approaches often suffer from significant limitations. Previous studies have frequently relied on small numbers of bacterial species, restricted datasets, single train-test splits, and the extraction of isolated colony crops rather than analyzing complete culture plates. There is a demonstrated need for a systematic, reproducible comparison of deep learning (DL) architectures capable of handling whole-plate images under consistent experimental conditions to support scalable and cost-effective bacterial identification.

Methodology

This study proposes a unified deep learning pipeline for multiclass bacterial species classification using complete, high-resolution culture-plate images. The methodology encompasses the following key components:

Dataset

The research utilized a balanced dataset comprising 950 high-resolution images representing 19 clinically relevant bacterial species derived from 151 clinical strains. The dataset was collected between August 2023 and May 2025 from respiratory, urinary, blood, and other clinical specimens. Images were acquired using a Micro-Chronos automated imaging system in a light-proof environment with dual-source lighting (transmitted and annular reflected) to capture colony boundaries and surface textures. The original resolution was 4056×3040 pixels, with exactly 50 images per species.

Preprocessing and Augmentation

A consistent preprocessing pipeline was applied to all images:

Conversion and Resizing: Images were converted to RGB and resized to 384×384 pixels using bilinear interpolation.
Normalization: Model-specific normalization was applied based on pretrained configurations (ImageNet parameters for CNNs; uniform parameters for ViT).
Data Augmentation: Stochastic augmentations (horizontal/vertical flipping, random rotation ±10°, translation ±5%, scaling 0.9–1.1, and color jittering) were applied exclusively to the training subsets of each fold. Validation sets remained unaltered.

Experimental Protocol

The study employed a stratified five-fold cross-validation strategy with a fixed random seed (42) to ensure identical data partitions across all models. This approach allowed for a rigorous comparison without the need for a separate held-out test set.

Comparative Models

Three distinct architectures were fine-tuned under identical optimization settings (AdamW optimizer, cross-entropy loss, 10 epochs, mixed-precision training):

ResNet50: A residual CNN utilizing bottleneck blocks to address degradation in deep networks.
DenseNet121: A CNN employing dense connectivity to reuse feature maps across layers.
ViT-B/16 (Vision Transformer): A transformer-based model that divides images into non-overlapping patches and utilizes global self-attention mechanisms.

All models were initialized with pretrained weights (ImageNet for CNNs; ImageNet-21K/1K for ViT) and modified with a task-specific classification head for 19 classes.

Key Contributions

The paper outlines the following primary contributions:

Whole-Plate Classification: Successful classification of 19 bacterial species using 950 complete culture-plate images, avoiding the need for prior extraction of individual colonies.
Unified Comparative Baseline: A direct, fair comparison of ResNet50, DenseNet121, and ViT-B/16 under identical preprocessing, augmentation, optimization, and cross-validation protocols.
Paradigm Comparison: An evaluation of three feature-learning paradigms: residual learning (ResNet), dense feature reuse (DenseNet), and global self-attention (ViT) for microbial identification.
Comprehensive Evaluation: Performance assessment using a wide range of metrics, including accuracy, macro-precision/recall/F1-score, Matthews Correlation Coefficient (MCC), Cohen's kappa, and detailed error analysis.

Results

The experimental results demonstrated a clear performance hierarchy among the models:

ViT-B/16 achieved the highest overall performance with a mean accuracy of 99.05 ± 0.86% and a macro F1-score of 99.05 ± 0.86%. It also achieved the highest MCC (0.9901) and Cohen's kappa (0.9900), with only 9 misclassifications across the aggregated out-of-fold predictions.
DenseNet121 ranked second, achieving 97.58 ± 1.69% accuracy and a macro F1-score of 97.56 ± 1.68%, with 23 total misclassifications.
ResNet50 achieved 94.21 ± 1.75% accuracy and a macro F1-score of 94.19 ± 1.75%, with 55 misclassifications.

Statistical Analysis:
Friedman tests indicated significant differences among the models ( $p=0.016$ ). Pairwise comparisons showed that both ViT-B/16 and DenseNet121 significantly outperformed ResNet50. However, the difference between ViT-B/16 and DenseNet121 was not statistically significant ( $p=0.108$ ) given the limited number of folds, though ViT-B/16 exhibited lower variability (SD 0.86% vs. 1.69%).

Error Analysis:
Misclassifications were concentrated in specific visually challenging pairs. Notably, the confusion between Acinetobacter baumannii and Serratia marcescens occurred across all models, suggesting overlapping visual characteristics. The confusion between Staphylococcus epidermidis and Corynebacterium striatum was prominent in CNNs but absent in ViT-B/16 errors, potentially indicating the transformer's ability to leverage global plate-level context.

Significance and Claims

The authors claim that this study provides a reproducible comparative baseline for bacterial image classification, demonstrating that deep learning models, particularly Vision Transformers, can achieve highly accurate species identification directly from whole-plate images without isolating individual colonies.

The study emphasizes that while ViT-B/16 achieved superior accuracy and stability, DenseNet121 remains a strong, computationally efficient alternative for scenarios with limited resources. The paper modestly concludes that while the results are promising, they are based on a single dataset under standardized conditions. The authors explicitly state that the findings motivate future research into strain-independent validation (to ensure generalization to unseen strains) and multi-centre validation across different imaging systems and laboratory environments. The study does not claim to solve all biological identification challenges but establishes a robust methodological framework for future comparative DL studies in microbiology.