I have successfully defended my dissertation on June 12th, 2019 at Department of Computer Science, University of North Carolina at Chapel Hill. My Ph.D research focus on Medical Image Analysis with Deep Learning from Sep 2015 to May 2019. In the 1st year (Aug 2014-May 2015) of my Ph.D, I work on probabilistic graphical model and RBM. From the 2nd year on, I moved to medical image analysis field. Specifically, I work on adversarial confidence learning with its application in medical image segmentation and synthesis. Another major research topic for my Ph.D defense is about low-contrast (blurry) boundary delineation. I also participate in other medical image analysis related projects.
Prior to UNC, I obtained my Master degree from University of Chinese Academy of Sciences in Computer Science where I studied text mining. My bachelor degree came from Northeastern University. I had an internship at United Imaging Intelligence in summer 2018, Institute of Deep Learning (IDL), Baidu Corporation in summer 2015, and an intership at NEUSOFT in summer 2010.
You can check my github if you are interested in some of my publications or projects. You can also contact me at ginobilinie.at.gmail.com.
My current main research focuses on adversarial machine learning, medical image segmentation, synthesis and classification problems. Beside, I take interest in implementing the large-scale parallel computing platform.
My recent work includes:
Adversarial Confidence Learning: focus on adversarial learning in supervised models, especially explore how to align the quantitative performance gain with visual perception performance improvement.
analyze roles of discriminator in the classic GANs and compare with those in supervised adversarial learning systems; simple application of adversarial learning in supervised models actually work as realistic regularization, which means the visual perception performance improve, but the quantitative performance does not improve accordingly (it even become worse in some cases); figure out that softmax probability usually pushes up the uncertainty and is thus not appropriate for representing the uncertainty; propose an adversarial confidence learning framework which aims at improving quantitative performance with confidence learning and retaining the visual perception performance gain at the same time, which mainly benefits from the point that the adversarial probability is the suitable uncertainty representation.
Medical image segmentation: mainly work on how to delineate the blurry boundaries, how to take advantage of large scale of unlabeled data and how to enhance the segmentation of difficult regions.
In particular, explore proper architectures to segment the blurry organs of an medical image; design effective (mainly adversarial confidence learning) algorithms to correctly include unlabeled data (SSL) to help the segmentation; design difficulty-aware attention mechanism (based on the relaxed adversarial confidence learning, and form a adversarial reweighting technique) for hard-to-segment region classification; explore basic network operations which adapts to medical image tasks; apply the developed algorithms to real clinic tasks, such as prostate and infant brain segmentation.
Medical image synthesis: design FCN and GAN based methods (arch and losses) to accurately and perceptually synthesis (reconstruct) medical images with context-aware mechanism; design difficult-region-aware attention mechanism to better model hard-to-predict regions (e.g., lesion/tumor region).
Medical image classification: design algorithms to effectively deal with multi-modality data to improve the classification tasks.
My previous work includes:
Mental Health Analysis: Mining depression and anxiety through analyzing the web browser history and user behavior with naive Bayes and SVM.
Personality Recognition: Recognizing personality traits through social media profile and posts analysis. We mainly focus on how to use numerous unlabeled data (semi-supervised regression) to train better and more stable models.
Bacteria Colony Analysis: Understand how colonies interact with each other using Convolutional RBM as analysis tool.