Pre-Trained Deep Convolutional Neural Network for Clostridioides Difficile Bacteria Cytotoxicity Classification Based on Fluorescence Images

• Main Authors: Andrzej Brodzicki, Joanna Jaworek-Korjakowska, Pawel Kleczek, Megan Garland, Matthew Bogyo
• Format: Article
• Language: English
• Published: MDPI AG 2020-11-01
• Series: Sensors
• Subjects: clostridioides difficile; fluorescence images; image analysis; classification; deep neural networks; convolutional neural networks
• Online Access: https://www.mdpi.com/1424-8220/20/23/6713

<i>Clostridioides difficile</i> infection (CDI) is an enteric bacterial disease that is increasing in incidence worldwide. Symptoms of CDI range from mild diarrhea to severe life-threatening inflammation of the colon. While antibiotics are standard-of-care treatments for CDI, they are also the biggest risk factor for development of CDI and recurrence. Therefore, novel therapies that successfully treat CDI and protect against recurrence are an unmet clinical need. Screening for novel drug leads is often tested by manual image analysis. The process is slow, tedious and is subject to human error and bias. So far, little work has focused on computer-aided screening for drug leads based on fluorescence images. Here, we propose a novel method to identify characteristic morphological changes in human fibroblast cells exposed to <i>C. difficile</i> toxins based on computer vision algorithms supported by deep learning methods. Classical image processing algorithms for the pre-processing stage are used together with an adjusted pre-trained deep convolutional neural network responsible for cell classification. In this study, we take advantage of transfer learning methodology by examining pre-trained VGG-19, ResNet50, Xception, and DenseNet121 convolutional neural network (CNN) models with adjusted, densely connected classifiers. We compare the obtained results with those of other machine learning algorithms and also visualize and interpret them. The proposed models have been evaluated on a dataset containing 369 images with 6112 cases. DenseNet121 achieved the highest results with a 93.5% accuracy, 92% sensitivity, and 95% specificity, respectively.