Yi Zheng

I am currently a Ph.D. candidate in Computer Science Department at Boston University, advised by Prof. Margrit Betke and Prof. Vijaya B. Kolachalama. I obtained my master degree in Electrical Engineering at University of Southern California. During my undergraduate study I've interned in Biomedical Imaging Lab led by Prof. Boqiang Liu at the Shandong University. I obtained my bachelor degree in Electrical Engineering from Shandong University, China.

Before Ph.D. study, I was an algorithm engineer in image quality team in General Electrics, Healthcare, where I worked on developing image quality tools & algorithms, conducting verification & validation of radiological imaging equipment.

I am interested in computer vision, representation learning, and deep learning, I currently focus on representation learning on high-resolution images for tissue phenotyping in computational pathology. It aims at characterizing objective and histopathologic features within gigapixel whole slide images (WSIs) for cancer diagnosis and the estimation of survival time in patients. Unlike CNNs/Vision Transformer which work on downsampled fixed size images, we are dealing with gigapixel images with variable sizes. We extract fine-grained features (such as stroma, tumor cells, lymphocytes), capture local clusters of cell-to-cell interactions (such as tumor cellularity), characterize macro-scale interactions between clusters of cells and their organization in tissue (such as the extent of tumorimmune localization in describing tumor-infiltrating versus tumor-distal lymphocytes), and finally figure out the overall intra-tumoral heterogeneity of the tissue microenvironment depicted at the slide-level of the WSI.

Email  /  CV  /  Google Scholar  /  Github

Publications

* indicates equal contribution

We present an attention-based fusion architecture that integrates a graph representation of pathology images with gene expression data and concomitantly learns from the fused information to predict patient-specific survival.

We present a Graph-Transformer (GT) based framework for processing pathology data, called GTP, that interprets morphological and spatial information at the WSI-level to predict disease grade.

Semantic-based Sentence Recognition (SSR) can efficiently understand the context between regions of text or between words in images by extracting sentences or paragraphs from images instead of only isolated text regions or words.

Our framework to analyzing microscopic- and WSI-level changes in renal biopsies attempts to mimic the pathologist and provides a regional and contextual estimation of IFTA. Such methods can assist clinicopathologic diagnosis.

Linguistically Aware Learning (LAL) scene text recognizer is a a bimodal framework that simultaneously utilizes visual and linguistic information to enhance scene text recognition performance.

UHT, short for UNet, Heatmap, and Textfill, uses a UNet to compute heatmaps for candidate text regions and a textfill algorithm to produce tight polygonal boundaries around each word in the candidate text.

We propose a scrape-by-location methodology to create food image datasets from Instagram posts. We applied our techniques to the millions of Instagram posts we collected across Kenya over a period of 20 days to give an example of the kind of analysis social scientists may conduct with our tools.

We developed a graph convolutional network that can efficiently process WSIs of hematoxylin and eosin (H&E)-stained NAT and tumor tissues from lung squamous cell carcinoma (LSCC) and lung adenocarcinoma (LUAD) cases.our analysis identified unique WSI-level signatures of NAT for LSCC and NAT for LUAD, which are distinct from those obtained on tumor WSIs. These findings underscore the importance of characterizing NAT samples in the context of their proximity to tumor subtype, and such phenotyping can facilitate studies on histologically precancerous lesions.

A deep learning framework known as a feature pyramid network (FPN) was implemented to classify digitized renal biopsies as class I or II DN. Our study identified several regions on the biopsy images as informative for prediction of class I vs II DN. Further analysis can elucidate the importance of various histopathological features of early stage DN.