School of Engineering Department of Computer Science and Engineering 101 Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: LIU, Runsheng / SENG Course: UROP1000, Summer Cell instance segmentation in cytology images holds significant importance for biological analysis and cancer screening; however, it remains a challenging task primarily due to two critical factors. Firstly, the presence of extensive overlapping cell clusters leads to ambiguous segmentation. Secondly, the existence of numerous adjacent cells with touching boundaries complicates the computer vision task, introducing further ambiguity in identifying real regions of cells. In particular, unification of two challenges makes accurate segmentation difficult to realize. In this work, we proposed a Gradient Anomaly Guided Network (GAGNet) to resolve the overlapping and touching challenges simultaneously. The datasets we plan to use are multiple datasets, because the aim of our method is to unify overlapping and touching problems and simultaneously resolve them. We design a strategy that filters and highlight two cases taking advantage of gradient anomaly. For touching case, we make use of direction difference to add attention on touching boundaries. For overlapping case, we obtain highlight maps with attentions from gradient anomaly, where deep neural network is following afterwards with the aim of obtaining precise overlapping area segmentation. After merging with complement ground truth map, together with direction difference map, overlapping and touching refinement takes place on original coarse segmentation mask. We are currently at the progress on designing the models and computation methods. Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: XIAO, Yicong / QFIN Course: UROP1100, Summer Histopathology is crucial for diagnosing diseases, but analyzing whole slide images (WSIs) manually is timeconsuming. With the development of WSI technology, there is a growing interest in developing automated methods to extract diagnostic information from WSIs, yet these methods usually suffer from the unbalance number of normal and tumor-indicative patches in a single WSI. This project aims to develop a stand-alone K-Means based pruning module that removes patches highly possible to be normal before passing a WSI into a downstream classification model. Completed experiments have show that the performance of the CLAMSB model integrated with this pruning model is better than the original model.
RkJQdWJsaXNoZXIy NDk5Njg=