Modelling Efficient and Robust Solutions for Microbiology Image Analysis Using Deep Learning

Abstract

Microscopic image analysis plays a crucial role in clinical microbiology laboratories for diagnostic purposes. Highly skilled microbiologists, also known as pathologists, are required to interpret vari- ous images, including Gram stain smears. These samples contain vital diagnostic information, such as identifying the presence and types of bacteria, evaluating specimen quality, and cell counting. However, manual interpretation of conventional glass microscopy slides remains a time-consuming, labour-intensive, and operator-dependent process. In high-volume pathology laboratories, implement- ing an artificial intelligent system could offer significant advantages by alleviating limitations faced by conventional pathology on a larger scale. Such a system would ensure enhanced accuracy, reduced workload for pathologists, and improved objectivity and efficiency. Consequently, this has motivated the research using data-driven techniques to develop automated interpretations of pathology images, particularly focusing on Gram stains. With the vast development and advancement in computer vision techniques, researchers have been able to explore the realm of Computer-Aided Diagnoses (CAD). The emergence of deep learning has revolutionised the analysis of pathology and medical images, moving away from traditional handcrafted features to leveraging the power of deep learning algorithms. Among these algorithms, Convolutional Neural Networks (CNNs) have demonstrated their ability to learn features from datasets, leading to enhanced performance and increased robustness of classifiers and detectors against variations Despite the extensive literature on pathology images, the automatic analysis of the Gram stain test using CNNs has not gained the same level of attention as other pathology tests such as breast cancer, lymphoma and colorectal cancer. It is exceedingly rare to find datasets relating to the very important Gram stain, and this data scarcity has likely hindered research on Gram stain automation and limited research in this area. This thesis aims to apply deep learning techniques to analyse pathology images, with a specific focus on Gram stain data. The aim is to discover novel approaches that can enhance the accuracy and efficiency of Gram stain analysis, bridging the gap in research and paving the way for advancements in this critical area. Initially, a CNN-based classifier was proposed for Gram-positive cocci bacteria subtypes in blood cultures. Throughout the study, the effect of downsampling, data augmentation, and image size on classification accuracy and speed was studied. To conduct these experiments, a novel dataset provided by Sullivan Nicolaides Pathology (SNP) consisting of three distinct bacteria subtypes, namely Staphylococcus, Enterococcus and Streptococcus were used. The sub-images were obtained from blood culture WSIs captured by the in-house SNP MicroLab using a ×63 objective without coverslips or oil immersion. The results show that a CNN-based classifier distinguishes between these bacteria subtypes with high classification accuracy. Secondly, existing CNN classification backbones operate under the assumption that all testing classes have been encountered during model training. However, in certain scenarios, it may be infeasible to collect all bacteria subtypes during the model training phase. CNNs are incapable of estimating their uncertainty, and they assume full knowledge of the world. To avoid misdiagnosis risk in the bacteria classification task, OpenGram a framework to open CNN classifier was proposed in this study that aims to tackle the problem of bacteria subtyping from an open-set perspective. Open-set recognition models can classify known instances and detect unknown samples of novel classes. OpenGram combines a CNN classifier with a Gaussian mixtures model to adapt to open-set classification. The results demonstrate OpenGram’s efficacy in accurately detecting unknown bacteria classes that were not encountered by the network during training, while maintaining the ability to classify known bacteria classes. Thirdly, most deep learning-based object detection methods rely on the availability of large sets of annotated training data, assuming that both training and testing data belong to the same feature space. However, these assumptions may not always hold true in real-world applications, particularly in the domain of pathology images. The process of collecting annotations for pathology images can be costly and labor-intensive. Additionally, testing supervised models on different distributions can degrade detector performance as these models might not be properly generalised to other domains. The objective was to tackle this lack of instance-level cell labels in Gram stain WSIs for the epithelial and leukocyte cell counting task. HybridGram, a framework with image translation and pseudo- labelling modules to completely avoid manual labelling on a new dataset was presented. The results demonstrate that HybridGram effectively bridges the performance gap between fully supervised and unsupervised models in this context.