Disease Progression Modeling

Disease progression models (DPM) aim to recover the trajectory of disease progression using a collection of short time series. A time series corresponds to multiple visits of a patient. At each visit, a patient is examined and various measurements (MRI, PET/SPECT, CSF, clinical scores) about the disease state are acquired. By combining these imaging/CSF/clinical biomarkers from all the patients, the disease progression trajectory can be retrieved. Our research is mainly focused on neurodegenerative diseases, such as Alzheimer's disease (AD), Parkinson's disease (PD) (IPMI 2023, TMI 2021, IPMI 2019).

Interpretable Machine Learning

Computer aided diagnosis using deep learning can achieve very high accuracy in classifying disease from normals in medical images. However, without explaining which part in the image contributes to the decision, the results are unreliable in clinical practice. We use the Shapley value, saliency maps, and other interpretable network architectures to make interpretable deep learning in computer aided diagnoisis of AD, PD, ASD using MRI/SPECT/fMRI (MedIA 2023, MICCAI 2022, BIBM 2021, MedIA 2021, MICCAI 2020 ).

Medical Image Segmentation

Traditional medical image segmentation is achieved by handcrafted algorithms targeted to a specific region. Nowadays, segmentation is achieved by training based algorithms (such as U-Net) that treat segmentation as a pixel-wise classification problem. We apply contrastive learning in semi-supervised medical image segmentation (TMI 2022).

Multiband Image Registration/Fusion

A hyperspectral image has a low spatial resolution and a high spectral resolution. A multispectral image, on the contrary, has a high spatial resolution but a low spectral resolution. Fusing these two types of images can lead to a both spatially and spectrally high-resolution image. Before fusion, they need to be registered. We propose a simultaneous registration and fusion algorithm to combine these two types of images (TGRS 2020).

Hyperspectral Image Unmixing

Hyperspectral images, despite their high spectral resolution that can characterize different materials, have a low spatial resolution (e.g. 16m per pixel). Hence, a pixel in a hyperspectral image typically contains multiple materials. How to find the spectra of these materials and their proportions in a pixel --- a.k.a. hyperspectral unmixing --- is a core problem in hyperspectral image analysis. We propose a Gaussian mixture model to unmix hyperspectral images (RSE 2020, TIP 2018, TIP 2016).