Battling Confounders of Deep Learning Models
Presence of confounding effects is the most critical challenge in applying deep learning techniques to medical applications. Confounders are extraneous variables that influence both input and output variables and thereby can easily distort the training and interpretation of deep learning models. While confounder control has been the center of discussion for traditional models, this topic is largely overlooked in the surge of deep learning applications as researchers put more attention on designing deeper and more powerful network architectures. To address this issue, we propose to learn features impartial to confounders (unbiased features) by adversarial training [CF-Net, Nature Communications][BR-Net][github], a generic layer correcting for metadata variables [Metadata Normalization, CVPR][github], and a visualization method that can disentangle the confounding effects falsely learned by the model. [Confounder-aware Visualization][github]
Translational Nueroimaging Study on the Impact of Alcohol Use
► [Alcohol use effects on adolescent brain development revealed by simultaneously removing confounding factors, identifying morphometric patterns, and classifying individuals, Scientific Reports, 2018]
Tools for Longitudinal MRI Analysis
Longitudinal neuroimaging studies have become increasingly prevalent these days. Longitudinal analysis of structural and functional organization of the brain, however, still relies on cross-sectional procedures (computational methods), which neglect intra-subject dependencies of longitudinal MRI data. We developed novel methods for characterizing macrostructural [LNE][LSSL][Longitudinal Pooling] and functional neurodevelopment [L-ICA][github] that reflects biologically plausible longitudinal effects. We then improved the analysis of longitudinal trajectories of connectivity patterns by studying the manifold of positive-definite cone (that underlies the connectivity data) incorporating theories of parallel-transport and Lie group action [Riemannian Geometry][github]. These methods were shown to have improved statistical power in detecting group differences in functional development of the brain.
Learning Underlying Geometry of Neuroimaging Data via Deep Generative Models
One central challenge in neuroimaging studies is that data often follow complex distributions in a high-dimensional image space. Therefore, learning the underlying low-dimensional latent space has been critical for successfully uncovering neuroscientific findings. Leveraging recent advances in deep learning, we have explored novel generative models to equip the latent space with the capability of feature disentanglement and characterizing multi-modal distribution. The resulting models can adapt to both supervised or unsupervised applications based on both structural and functional MRI data, such as to understand how the brain structures change with age in both healthy aging and in neurodegenerative diseases [LSSL][VAE-R][github], and to discover major patterns of functional brain connectivity [tGM-VAE][github].
Interpretable Deep Learning Models for Neuroimaging Applications
When deep learning methods are applied to neuroimaging and computational neuroscience applications, it is not only important to enable prediction of disease outcomes, but also to understand the underlying reasons why each subject is classified with any specific disease. Such an interpretation contributes to a mechanistic understanding of the diseases and may help clinicians design new therapeutic procedures. To this end, we conducted several research projects on the interpretability and explainability of deep learning methods applied to neuroimages [DeepVisualization][VAE-R for T1-w MRI][github] [ST-GCN for rs-fMRI][github]. We proposed 3D Convolutional Neural Networks and exploited their model parameters to tailor the end-to-end architecture for the diagnosis of different diseases from MRIs. Based on the learned models, we identified disease biomarkers and validated the results by exploring the importance of brain regions (or voxels) [Visualizing sex differences during pre-adolescence] to the model prediction and by relating the findings to the clinical literature.
Endoscopic Video 3D Reconstruction and Registration with CT
The clinical problem we want to tackle here is to transfer the tumor information from a 2D endoscopic movie frame into the 3D CT space for radiation treatment planning. The solution is to first reconstruct a 3D surface model from the video and then register that surface to the CT image. We developed methods for fusing single-frame reconstructions into a complete surface based on physical and statistical models [Geometry Fusion][Anisotropic Stiffness Learning][Joint Disparity-estimation/Registration]. The reconstruction surface was then mapped to the CT space based on spectral and physical models [Thin Shell Demons] [Spectral Graph Theory].
2D/3D Registration for Abdomen Radiation Treatment Planning
2D/3D registration is often used in Image-Guided Radiation Therapy (IGRT) for tracking target motion during treatment delivery. A challenge in disease sites subject to respiratory motion is that organ deformation may occur, thus requiring incorporation of deformation in the registration process. An improved metric-learning algorithm was designed for this purpose. We were the first ones to study this problem in the abdomen. [Local Metric Learning]