Skip to main content

Research

Connecting imaging physics and machine learning with structure and physical properties

Table of Contents

    From imaging measurements to structural and physical understanding

    SMILE Lab studies how to recover images from limited, noisy, or motion-corrupted measurements, and how to use them to understand anatomy, development, and physical properties. We bring imaging physics and machine learning together across reconstruction, quantitative analysis, image restoration, and 3D modeling.

    Across these areas, we ask how learned priors can remain faithful to measurements, how methods can depend less on paired training data, and how models can adapt to different subjects and acquisition conditions. The sections below connect research questions with methods and results. Papers, code, project pages, and posters are collected in Publication.

    🔬

    Physics-Guided Reconstruction

    Continuous representations & priors

    Combine neural representations, imaging physics, and diffusion priors for undersampled reconstruction, motion correction, artifact reduction, and geometry adaptation.

    Explore this direction
    🧠

    Developmental Imaging & Analysis

    Continuous atlases & quantitative measures

    Study fetal brain and body atlases, adult brain tissue, and quantitative methods that connect anatomical variation across subjects and organs.

    Explore this direction
    ✨

    Denoising & Enhancement

    Self-supervision & degradation modeling

    Study single-image denoising, low-field MRI enhancement, and restoration under unknown degradation without paired reference images.

    Explore this direction
    🦴

    Anatomy & Physical Understanding

    From surface shape to internal structure

    Connect anatomy-aware parametric models, surgical shape prediction, and real-world object density fields through representations of internal structure.

    Explore this direction
    51
    Research publications
    13
    CCF-A conference papers
    4
    Core research areas
    2
    NSFC General Program grants
    🎯

    Start with the imaging problem

    Design around undersampling, motion, noise, and acquisition geometry, keeping actual observations in the reconstruction process.

    🧩

    Make data requirements explicit

    Distinguish self-supervision, scan-specific optimization, and pretrained priors, with attention to each method's assumptions and scope.

    💡

    Share methods and resources

    Bring together released code, atlases, datasets, and project pages to support reproduction, comparison, and further research.


    Content updated: October 2026

    Research Directions and Contributions

    Research questions, methodological approaches, and representative work

    Direction 1

    Continuous Representations and Physics-Guided Reconstruction

    How can we recover reliable medical images from incomplete measurements affected by motion or artifacts? We combine implicit neural representations, imaging physics, and generative priors for MRI, CT, and X-ray reconstruction. Beyond visual quality, we focus on agreement with measured data and adaptability to changes in sampling, acquisition geometry, and hardware.

    Continuous Representations and Physics-Guided Reconstruction
    Direction 2

    Developmental Imaging and Quantitative Analysis

    From fetal development to adult brain structure, we study how medical images can provide comparable, quantitative anatomical information. Our work centers on fetal MRI reconstruction, continuous developmental atlases, and quantitative brain analysis, with extensions to fetal body atlases, pathology, and clinical imaging. We connect image reconstruction with biomedical questions while preserving both population trends and individual variation.

    Developmental Imaging and Quantitative Analysis
    Direction 3

    Self-Supervised Denoising and Image Enhancement

    Low signal-to-noise ratio, limited resolution, and unknown intensity mappings affect both image detail and quantitative information. We study restoration without paired reference images through two complementary approaches: self-supervision from individual noisy images, and pretrained priors combined with degradation modeling. Our work spans MRI, fluorescence and electron microscopy, and blind inverse problems with unknown dynamic range compression.

    Self-Supervised Denoising and Image Enhancement
    Direction 4

    Anatomical Modeling and 3D Physical Understanding

    A 3D model should describe more than surface shape: it should also capture internal structures and their relationships. Using CT, MRI, and X-rays, we study joint bone–soft-tissue modeling, parametric hand representations, and facial shape prediction after orthognathic surgery. Extending this focus from appearance to internal structure, we also study density fields of real-world objects for physical property estimation and interaction.

    Anatomical Modeling and 3D Physical Understanding