A Neural Network Approach for Yeast Colony Counting in Petri Plate Images

This work presents a methodology for identifying and counting colonies of a specific type of yeast on images of Petri plates colonized by different types of fungi. Images of Petri plates are commonly used in biological research to analyze the growth of microorganisms. It usually involves the manual counting of the number of colonies, which can be time-consuming and prone to errors, especially when multiple types of microorganisms are present. The proposed approach utilizes an Artificial Neural Network (ANN), combined with image processing techniques, to Filter Out unwanted colonies and other artifacts from the images (for instance, light reflections and plate edges). The mask generated by the ANN is then used to the automatic counting of the colonies of interest. While artificial intelligence approaches typically require large amounts of training data, the here proposed methodology is designed to minimize the amount of labeled data required for training, making the approach more practical for laboratory settings.

An application is presented in the context of counting colonies of <em>Candida krusei</em> on Petri plates also colonized by <em>Candida glabrata</em>. The image of the Petri plates were captured using a low-cost setup consisting of a smartphone camera and a simple lightbox to ensure uniform lighting conditions. The ANN filter is designed as a Multilayer Perceptron (MLP) trained to classify small disk patches extracted from the original images, which were resized to match the patch size to the diameter of the smallest colonies. The input to the MLP consists of the gray level values of the pixels in the patch, while the output is a real number between 0 and 1 indicating the likelihood that the pixel at the center of the patch belongs to the target colony type. An early stop technique, balanced training samples, and a sample-rotation data augmentation strategy were employed to enhance the training process and improve the generalization capability of the MLP. By applying the trained MLP to each pixel in the image, a probability map (mask) is generated. The generated mask contains local maxima corresponding to the centers of the colonies of interest. A peak detection algorithm is then applied to the mask to identify and count the colonies.

A preliminary evaluation of the proposed methodology was conducted using two Petri images (one for training and another for testing). The training image was manually annotated to create a ground truth mask for training the MLP identification of the <em>Candida krusei</em>. The performance of the ANN filter was assessed by comparing the counts obtained from the generated mask with manual counts performed by experts. The ANN approach achieved a recall of 94.8%, precision of 97.9% and an F1-score of 96.3%, indicating high accuracy in detecting and counting the target yeast colonies in the test set.

As a work in progress, the current stage of development focuses on expanding the model to process RGB input patches, enabling the ANN filter to incorporate color information during training and inference. It is expected the improvement of its ability to differentiate multiple colony types, while handling greater visual variability in colony appearance. Future work also includes expanding the methodology validation with a larger and more diverse set of colony images to improve generalization across different colony morphologies. Additional improvements will involve experimenting with alternative ANN architectures designed for multi-channel inputs, optimizing filter inference time, and refining the peak detection algorithm to improve counting accuracy in cases where colonies overlap or present size variability, ultimately strengthening the robustness of the system.