I am a PhD research student at Machine Learning and Computer Vision Research Group (MaVi) and Visual Information Lab (VIL), University of Bristol, where I am supervised by Dr Alberto Gambaruto and Dr Tilo Burghardt. We work on the swallowing project CTAR-SwiFt.
My part of work focuses on Medical Image Processing using Deep Learning and Computational Fluids Dynamics using Meshless method/Physics-Informed Neural Network.
News
- 17/06/2024: I began a six-month internship as an Applied Scientist at Amazon, London, in Prime Video & Amazon MGM Studios.
- 18/04/2024: My 3-Minute Thesis “How AI helps doctors diagnose swallowing difficulty” was selected as a Semi-Finalist in the 3MT competition at the University of Bristol.
- 03/03/2024: Our “RBF-PINN” paper was accepted by the International Conference on Learning Representations (ICLR), AI4DifferentialEquations in Science workshop.
- 19/02/2024: Our research proposal “AI aided diagnosis of Dysphagia in Elderly Population” was chosen as one of the Finalists in the TakeAIM competition by the Smith Institute.
Projects
Tea Classification - 2024
Chengxi Zeng, Ted Littledale, Yorkshire Tea (Bettys & Taylors of Harrogate)
[WEB APP]
We developed a tea classification Web App for Yorkshire Tea (Bettys & Taylors of Harrogate). The app is able to classify tea colors from any cups. My main responsibility was to develop image processing code for the tea element and extract the main tone by applying various edge detection and outlier removal techniques. Additionally, I employed the segmentation foundation model Segment Anything on a small tea dataset I created. The model accurately segmented the tea element and was converted to an ONNX file that can be deployed to multiple platforms. Prior to UK national tea day, the app successfully attracted thousands of users in a few weeks.
RBF-PINN: Non-Fourier Positional Embedding in Physics-Informed Neural Networks - 2024
International Conference on Learning Representations (ICLR 2024) , AI4DifferentialEquations in Science Workshop, Accepted
Chengxi Zeng, Tilo Burghardt, Alberto Gambaruto
[Github] [arXiv]
While many recent Physics-Informed Neural Networks (PINNs) variants have had considerable success in solving Partial Differential Equations, the empirical benefits of feature mapping drawn from the broader Neural Representations research have been largely overlooked. We highlight the limitations of widely used Fourier-based feature mapping in certain situations and suggest the use of the conditionally positive definite Radial Basis Function. The empirical findings demonstrate the effectiveness of our approach across a variety of forward and inverse problem cases. Our method can be seamlessly integrated into coordinate-based input neural networks and contribute to the wider field of PINNs research.
Training dynamics in Physics-Informed Neural Networks with feature mapping - 2024
Preprint, Under Review
Chengxi Zeng, Tilo Burghardt, Alberto Gambaruto
[Github] [arXiv]
Physics-Informed Neural Networks (PINNs) have emerged as an iconic machine learning approach for solving Partial Differential Equations (PDEs). Although its variants have achieved significant progress, the empirical success of utilising feature mapping from the wider Implicit Neural Representations studies has been substantially neglected. We investigate the training dynamics of PINNs with a feature mapping layer via the limiting Conjugate Kernel and Neural Tangent Kernel, which sheds light on the convergence and generalisation of the model. We also show the inadequacy of commonly used Fourier-based feature mapping in some scenarios and propose the conditional positive definite Radial Basis Function as a better alternative. The empirical results reveal the efficacy of our method in diverse forward and inverse problem sets. This simple technique can be easily implemented in coordinate input networks and benefits the broad PINNs research.
Video-SwinUNet: Spatio-Temporal deep learning framework for VFSS instance segmentation - 2023
IEEE International Conference on Image Processing (ICIP 2023), Accepted
Chengxi Zeng, Xinyu Yang, David Smithard, Majid Mirmehdi, Alberto Gambaruto, Tilo Burghardt
[Github] [arXiv]
This paper presents a deep learning framework for medical video segmentation. Our proposed framework explicitly extracts features from neighbouring frames across the temporal dimension and incorporates them with a novel temporal feature blender which then tokenises the high-level Spatio-temporal feature to a strong global feature encoder Swin Transformer. Our model outperforms other approaches by a significant margin and improves the segmentation benchmarks on the VFSS2022 dataset, achieving a dice coefficient of 0.8986/0.8186 for Part1/Part2 data. Our studies have also shown the efficacy of the temporal feature blending scheme and the transferability of the framework.
Lithology Document Analysis - 2022
Chengxi Zeng, Arion.ai, CGG
We developed a deep learning pipeline that processes long lithology tracks to multi-page images, a fast bounding box detection algorithm is employed and calibrated. Segmentation models are used to extract the curves and hence the numerical data is restored. The pipeline is tested on 1000+ documents and can process 100+ Page/Sec
Video-TransUNet: Temporally Blended Vision Transformer for CT VFSS Instance Segmentation - 2022
SPIE International Conference on Machine Vision (ICMV 2022), Accepted, Best Oral Presentation
Chengxi Zeng, Xinyu Yang, Majid Mirmehdi, Alberto Gambaruto, Tilo Burghardt
[Github] [arXiv]
We propose Video-TransUNet, a deep architecture for instance segmentation in medical CT videos constructed by integrating temporal feature blending into the TransUNet deep learning framework. In particular, our approach amalgamates strong frame representation via a ResNet CNN backbone, multi-frame feature blending via a Temporal Context Module (TCM), non-local attention via a Vision Transformer, and reconstructive capabilities for multiple targets via a UNet-based convolutional-deconvolutional architecture with multiple heads.
Portable AED Delivery Emergency Drone (Patented) - 2022
Chengxi Zeng, Yuhang Ming, Mengxun Bai, Ruobing Li
We are developing an Emergency drone that delivers medical kits and portable AED which would save lives in jammed cities and remote areas. This project collaborates with Hangzhou Municipal Health Commission and the drone will work with local Emergency Response Unit.
Elastic Transformation Detection - 2021
Chengxi Zeng, Haytham Technology Ltd
We utilised CNN and other image processing techniques to non-destructively detect Elastic Transformation on steel/composite materials due to repetitive work load in extreme environments.
Solving Partial Differential Equations using Radial Basis Functions - 2021
Chengxi Zeng, Andreas Michael, Celia Tugores-Bonilla, Natasha Moore, Jules Boisser, Alberto Gambaruto
We discretise standard Partial Differential Equations(PDEs) using Radial Basis Functions, aiming to assess their performance compared to the analytical results and alternative discretisation methods.
Pedestrian Level wind assessment with CFD simulation - 2019
Chengxi Zeng, Alberto Gambaruto
We modelled Bristol city centre area and applied CFD simulation with focus on high wind speed area and presented the advantage of trees to mitigate wind using Porous Media approach.