Abstract
Microscopic cell image analysis plays a crucial role in understanding biological systems and facilitating disease diagnosis. Traditional image processing approaches typically rely on hand-crafted feature extraction and are designed for specific scenarios. However, real-world applications present a wide variety of variations in microscopic images, influenced by factors such as cell types, sample pre-processing techniques, and illumination conditions during image acquisition. While customized methods can be tailored to specific requirements, they may have limitations or constraints that result in suboptimal performance in other situations, posing challenges to meeting the demands of advanced healthcare automation. Fortunately, the emergence of deep learning technology has brought about substantial advancements in computer vision, opening up new possibilities for cell image analysis. This dissertation investigates deep learning-based approaches for microscopic cell image analysis.To start with, the study focuses on one of the common applications, namely cell classification. Deep learning is a powerful tool for object recognition. However, the lack of programming experience makes it difficult for novice users to apply this technology. To lower the barrier for clinical users to implement deep learning methods in single-cell image recognition, an out-of-the-box software called AIMIC (artificial intelligence-based microscopic image classifier) was developed. The platform is equipped with state-of-the-art deep learning-based classifiers and a data pre-processing pipeline. The users can apply deep learning technology with AIMIC in a fully code-free manner. Furthermore, the built-in networks were evaluated on four benchmark cell image datasets to assist entry-level practitioners in selecting a suitable classifier.
Then, a new interactive dual-network framework was developed to enhance accuracy and efficiency in cell counting. In this framework, one deep learning network (counter) is trained to regress a density map from a given microscope image. The number of cells in that image can be estimated by performing integration over the regressed density map. Another network (ground truth generator) is employed to generate suitable ground truth for dynamic supervision. The optimal model is obtained through an interactive process by iteratively and jointly training the counter and the ground truth generator. Additionally, a hierarchical multi-scale attention-based architecture was constructed to act as the counter for high-quality density map regression. Evaluation experiments on multiple public cell counting datasets demonstrated the superiority of the proposed method.
Lastly, the cell detection application was investigated. The automated detection of cells remains a challenging task due to the presence of inhomogeneous background textures, substantial inter-cell occlusions in dense regions, and significant variations in cell shapes and sizes. This issue was addressed by training a well-designed network to produce an inverse distance transform-based map and then extracting cell instances using local maxima guided by auxiliary counting. The inverse distance transform-based map can effectively highlight individual cell instances, while the cell count can help remove false responses, thereby enhancing detection accuracy. The proposed detection method demonstrated its robustness and universality by achieving superior detection performance in various scenarios.
| Date of Award | 21 Aug 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Jun LIU (Supervisor) |