Computer vision in histopathology image analysis: preprocessing and classification

In recent decades, the progress in computer vision techniques, particularly those based on deep learning, has significantly enhanced the field of digitalised medical image analysis. Concurrently, advancements in image scanning technology have propelled the widespread adoption of Whole-Slide Imaging (WSI) in clinical pathology. However, the direct implementation of deep learning methods on WSI images encounters several challenges, including variances in stain colour between images, class imbalances, and resource-intensive training pipelines. The first part of this thesis focuses on developing novel stain normalisation methods using Generative Adversarial Networks (GANs). Specifically, we developed two GAN models: the first model focuses on enhancing the texture features of the stain-normalised images, while the second model uses a semi-supervised learning framework that incorporates additional images with heterogeneous colours to further improve the quality of stain-normalised images. The second topic of this thesis focuses on data-efficient learning for WSI classification. WSIs are massive and challenging to process in their entirety using deep learning models, thus they are typically divided into thousands or tens of thousands of smaller-sized patches for processing. In fact, WSI classification can be conducted with a small set of key instances. To find these instances, we introduce a dataset distillation framework tailored to WSI analysis. In the third chapter, we focus on class imbalance learning which is a commonly known issue in WSI datasets. Here, we proposed a method to enhance the feature learning capacity of a Convolutional Neural Network through graph attention networks and contrastive learning. More- over, we also focus on long-tailed learning, which is a more challenging task than imbalanced learning due to a larger number of class labels. To address this issue, we proposed a method that balances parameter contributions to different classes, demonstrating that this approach can alleviate long-tailed classification problems. To sum up, the methods proposed in this thesis address challenges throughout the WSI analysis pipeline as well as some difficult long-tailed classification tasks in the general image domain. The results obtained from these methods demonstrate their potential to improve the processing efficiency of deep learning models and bring us closer to their use in real-world clinical scenarios.