Chen Jin

We introduced an approach for learning to downscale high-resolution images for segmentation tasks. The main motivation was to adapt the sampling budget to the difficulty of segmented pixels or regions. We showed that learning the spatially varying downsampling strategy jointly with segmentation offered advantages in segmenting large images with a limited computational budget.

As the leading author, I was fully responsible for development. This work was published at ICLR 2022.

Segmenting ultra-high resolution images often needs empirical decisions on the trade-off patch configuration between field-of-view (FoV) — spatial coverage — and the image resolution. We introduced the foveation module, a jointly learnable data loader which, for a given ultra-high resolution image, adaptively chooses the appropriate configuration (FoV or resolution trade-off) of the input patch to feed to the downstream segmentation model at each spatial location of the image.

The work was published at MICCAI 2020, and I was the leading author overseeing the development.

On top of prior foveation work at MICCAI 2020, in this extended version, we further introduced the more computational efficient hard-gated categorical sampling of FoV-resolution patch configurations at each location and provided two differentiable solutions. We demonstrated the generical applicability of three vision benchmarks, including Cityscapes autopilot images, DeepGlobe aerial images, and Gleason2019 histopathology images.

As the leading author overseeing the entire development, I published this work in Arxiv.

ROLE AND RESPONSIBILITIES
• Developed a self-supervised method for MR histology text representations learning, aiming for cheap MRI single-modal inferences.
• Created cross-modal-scale MRI histology image generation and mapping (upscaled 1,000 times) for explainable AI.
• Developed an interactive web UI based on a Gradio API to showcase registered cross-modality image patches.
• Disentangled inter-reader variability in segmentation and co-authored the MICCAI/NeurIPS 2020 publication.

ROLE AND RESPONSIBILITIES

• Supervised a PhD student on deep multiple instance learning (MIL) for large image classification and submission to journals.
• Supervised two MSc projects performing CNN-based segmentation and detection of histology images.
• Developed and supervised two MSc projects on super-resolution with CNN and GAN-based methods.

ROLE AND RESPONSIBILITIES

• Developed a weak supervised iterative convolutional net based on filter banks, AdaBoost, and auto-context that improves segmented connectivity of thin linear fracture. It was presented at conferences in 2016 and displayed as a poster in 2018.
• Conducted CT and SEM image registration, multiscale data fusion, and guided 3D pore structure reconstruction via multiple-point statistics method, conditional to neighboring "patch."
• Developed an automated patch-based rock pattern classification procedure based on random forest.
• Conducted modeling and numerical simulation of a macroscale reservoir multiphase flow presented to the industry.

ROLE AND RESPONSIBILITIES

• Wrote R scripts to merge pixel annotations from six experts into one.
• Developed and trained deep learning models for the segmentation and classification of prostate cancer Gleason grading on ultra-high-resolution histopathological tissue microarray (TMA) images.
• Developed evaluation codes to run on the test dataset as per the task request and submitted them to the competition. Entered the top ten on the leaderboard.

ROLE AND RESPONSIBILITY

• Developed and trained a segmentation model for ultra-high-resolution histology image segmentation and classification.
• Handled the preprocessing of ultra-high-resolution histology images into paired patches suitable for a deep learning model.
• Developed the evaluation code and packaged it in a Docker container for submission.

ROLE AND RESPONSIBILITY

• Included remote sensor image data cleaning and pre-processing.
• Deployed various segmentation models to balance performance efficiency, such as U-Net, MobileNet, and HR-Net.
• Implemented various optimization tricks to improve performance, such as augmentation and regularization.
• Developed the testing code and package as per the organizer's request for testing.
• Entered the final round of competition (top 10%).

Performed simulation of a two-phase (water/oil) reservoir. The models have regular and irregular shapes with wells at opposite corners to simulate production in a quarter-five-spot pattern. Providing reservoir properties, history matching, and water/oil displacement is simulated with Eclipse.