Noise aware preprocessing improves automated microbial colony enumeration in real world images

Introduction: Microbial colony enumeration is central to microbiological testing, including microbial limit testing used in pharmaceutical and consumer safety assessment. Although automated colony counting systems have been proposed, most rely on laboratory-grade images and perform poorly when applied to real world data. In practice, images of agar plates often contain uneven illumination, shadows, blur, and overlapping colonies, leading to substantial accuracy loss. This study addresses the lack of robustness of existing systems by introducing a noise aware preprocessing pipeline designed to improve colony enumeration under suboptimal imaging conditions.

Introduction: A web based machine learning pipeline for automated microbial colony enumeration was developed. The system consists of two stages: (1) image artifact correction and (2) colony detection. Artifact correction is performed using a GAN trained to remove common image distortions, including Gaussian noise, blur, and shadow artifacts. A noise type classification module is jointly trained with the denoising network to guide artifact correction. The processed images are subsequently passed to an object detection model based on a YOLO-architecture to localize and count colonies.

All models were trained using a batch size of 512 for 80 epochs on a GPU enabled Google-Colab environment, with a total training time of approximately 8 hours. Performance was evaluated through ablation studies comparing models without artifact correction, with denoising only, and with both denoising and noise type classification. Training and testing was performed on a publicly available Neurosys MLT Dataset consisting of 1,862 agar plate images across five microorganism species.

The baseline detection model achieved an accuracy of 85.7% on the full dataset and 77.2% on hard samples containing significant image artifacts. Incorporating the denoising module increased accuracy to 91.1% and 86.6%, respectively. The full pipeline, including both denoising and noise type classification, achieved an accuracy of 94.5% on the full dataset and 91.6% on hard samples, demonstrating consistent performance under non-ideal imaging conditions.

Noise aware preprocessing substantially improves the robustness of automated microbial colony enumeration in real-world images. These findings highlight the importance of explicitly addressing image artifacts in biomedical image-analysis pipelines, especially in the context of colony enumeration.

Keywords: Microbial Limit Testing, Colony Enumeration, Machine Learning, Image Artifact Correction