Search

Problem: Recently, deep convolutional neural networks have greatly improved our ability to develop robust image analysis methods, especially those tailored to biomedical scenarios. Despite important advances in computer vision, it is usually challenging to achieve the same success on specific biomedical datasets due to certain specific challenges, such as limited size datasets, expensive annotations, high-resolution information, multiple discernible biomedical objects, and lack of interpretability. Therefore, it is still in high demand to investigate automatic image analysis methods with high interpretability and accuracy to transform high-resolution biomedical image data into meaningful quantitative information with weak annotations and limited availability of training data. Methodology: Considering the challenges of biomedical datasets and the corresponding technical limitations, this thesis aims to incorporate global context-aware information to enable automated image analysis systems to provide reliable diagnostics in clinical applications. The thesis first introduces a hybrid method equipped with U-Net and a global probabilistic model for the segmentation of dense cell populations. This model jointly exploits shape priors and global cellular interactions to compensate for the limitations of standard deep networks. Next, this thesis focuses on attention-based, interpretable fine-grained classification methods for biomedical images that is capable of jointly capturing local high-resolution detail and global background information. Outcomes: In this thesis, a number of fine-grained feature learning methods were proposed for high-resolution biomedical images. The original contributions are as follows: (1) A global probabilistic model for dense cell nucleus segmentation. (2) A gated attention mechanism for simultaneous localisation of multiple discriminative instances. (3) A multi-task learning scheme for fine-grained feature learning. (4) Two new clinical datasets for biomedical image classification.