New Deep Learning Methods for Biomedical Image Segmentation and Related Problems

Accurately identifying objects in biomedical images plays a crucial role in biomedical image analysis, as it enables various downstream tasks such as cell behavior analysis and cancer diagnosis. However, this task poses inherent challenges due to the distinctive characteristics and complexities inherent in biomedical images.

Existing methods in biomedical image segmentation can be broadly categorized into two approaches: region-based methods and pixel-based methods. Region-based methods often rely on bounding box representations, which may limit their generality and make it difficult to accurately capture irregular or non-box-like shapes. On the other hand, pixel-based methods assign labels to individual pixels, allowing for more precise segmentation. However, pixel-based methods face challenges related to the clustering property when converting instance labels to pixel-level labels.

In my research, I have focused on improving the accuracy of biomedical image segmentation by incorporating prior knowledge, utilizing geometric features and proposing hierarchical models. The key contributions of my work can be summarized as follows: (1) Incorporating temporal/spatial instance consistency for video and 3D instance segmentation: Generating instance candidates from semantic segmentation masks and selecting an optimal subset of instance candidates by incorporating temporal/spatial instance consistency. (2) Incorporating geometric information for cell segmentation: Decomposing cell masks into boundary representations with explicit geometric features for classification as positive or negative samples. (3) Proposing a hierarchical decoder in Encoder-decoder deep learning models: Introducing a universal cascade decoder for biomedical image segmentation to fully leverage multi-scale features.

Furthermore, in addition to addressing image segmentation challenges, I have also devoted my research efforts to collaborating with researchers from biomedical research labs to solve fundamental biological research. Specifically, I introduced a new tracking framework that can segment and track Pseudomonas aeruginosa effectively for the analysis of single-cell behavior. I introduced an integrated detector-tracker that minimizes the impact of detection errors on the final tracking results. I developed a registration framework that can accurately register the wing disc pouch across image sequences, accommodating intricate deformations and challenging intensity distortions.