School of Engineering Department of Computer Science and Engineering 94 Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LIU Ang / MATH-CS Course: UROP 1100, Fall In this UROP project, which is focusing on virtual staining in the field of medical computer vision. Our group aimed to build up a deep learning based model that could perform virtual staining with the given WSIs. We plan to use the ACROBAT 2023 Grand Challenge dataset (https://acrobat.grand-challenge.org/), which consists of Whole Slide Images (WSIs) of different staining methods such as HE, as part of the training data of the model. In this project, I was assigned to do the preprocessing job of this dataset. In order to make the dataset usable for training, I decided to use VALIS (https://pypi.org/project/valis-wsi/) for image registration, and CLAM(https://github.com/mahmoodlab/CLAM) for segmentation to prepare the dataset. The registration is necessary for the large images such as WSIs, since we have to reduce the misalignment between the WSIs in the same labeled group, if the images have a lot of misalignment, the model may learn something useless or bad for generating the expected results. This project experience gives me a valuable insight into virtual staining and how to perform similar works in the future research. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LIU Runsheng / COMP Course: UROP 1100, Fall UROP 2100, Spring Instance segmentation plays a vital role in the detailed morphological quantification of biological entities such as tissues and cells, enabling precise identification and delineation of different structures. Current methods often address the challenges of touching, overlapping or crossing instances through individual modeling, while neglecting the intrinsic interrelation between these conditions. In this work, we propose a Gradient Anomaly-aware Medical Instance Segmentation approach GAInS, which leverages instance gradient information to perceive local gradient anomaly regions and models the spatial relationship between instances, thus refining local region segmentation. Specifically, GAInS is firstly built on a Gradient Anomaly Mapping Module (GAMM), which encodes the radial fields of instances through window sliding to obtain instance gradient anomaly maps. To efficiently refine boundaries and regions with high degree of gradient anomaly, we propose an Adaptive Local Refinement Module (ALRM) with a gradient anomaly-aware loss function. Extensive comparison and ablation experiments in three medical scenarios demonstrate that our proposed GAInS outperforms other state-of-the-art (SOTA) instance segmentation methods.
RkJQdWJsaXNoZXIy NDk5Njg=