New Algorithms and Deep Learning Methods for Object Detection and Segmentation for Biomedical Image Analysis

Machine learning and deep learning methodologies make it possible to achieve high-throughput and high-quality analyses of large volume medical and biological data. Data-driven biomedical image analysis has drawn increasing attention and interest in recent years.

Image analysis often starts with object detection and segmentation problems. Achieving good accuracy is probably the most important goal when solving detection and segmentation problems. More labeled data often helps train a better model. However, manual annotation is time-consuming and expensive to obtain, and annotation of biomedical images is commonly done most effectively by biomedical experts. Thus, it is often not realistic to expect large amounts of well-labeled data available for every deep learning based biomedical image analysis application. Utilizing unlabeled data together with labeled data has the potential to train a better model, but unlabeled data often requires certain assumptions and careful designs to be effectively utilized. The main goal of using machine/deep learning for solving detection and segmentation problems in biomedical image analysis can be summarized as obtaining good accuracy on model-unseen testing images without the needs of large amounts of labeled training data.

For achieving the above-mentioned goal, there are several main focuses in the current research field: (A) better model architectures, (B) better model training methods, (C) regularization techniques, (D) ways of utilizing unlabeled data, and (E) ways of utilizing prior knowledge. In this dissertation, we present new methods and algorithms that are related to these research focuses. More specifically, we present two new methods for utilizing unlabeled data for model training: (1) deep adversarial networks for biomedical image segmentation utilizing unannotated images, and (2) a new deep learning method using algorithm-generated pseudo-annotations. Furthermore, we present two new methods on utilizing prior knowledge for segmentation problems: (1) a new decompose-and-integrate learning method, and (2) a segmentation proposal generation method for H\&E stained histology images. A new data augmentation technique based on super-pixels is then presented as a new regularization technique for training biomedical image segmentation networks. Finally, on model architectures and training methods development, a coarse-to-fine fully convolutional network is presented for learning segmentation knowledge in a coarse-to-fine and simple-to-complex manner. In addition, we further present some algorithms we developed for ants tracking (building ants trajectories) and ants trajectory analysis.