Embedded Files
My research centers on inverse problems—inferring causal factors (e.g., initial conditions, model parameters) from limited observations—across medical imaging and computational biology. These problems are often ill-posed, meaning data alone is insufficient to guarantee a unique solution. To overcome this, I develop machine learning (ML) and deep learning (DL) algorithms capable of extracting relevant structures from noisy or incomplete measurements at remarkable speed and accuracy.
Key Focus Areas & Contributions
Medical Imaging
Deep Learning Reconstruction: Developed techniques to reconstruct high-quality B-mode images from sub-sampled radio-frequency (RF) data. This reduces hardware demands while preserving diagnostic clarity.
Adaptive & Compressive Beamforming:
Pioneered deep-learning-based beamformers that adapt to varying sensor configurations, enabling high-resolution ultrasound imaging with fewer measurements.
Switchable & Tunable Deep Beamforming:
Created a single model that can output multiple imaging “styles” (e.g., deconvolution, speckle removal) simply by changing a lightweight parameter, allowing real-time customization.
Unsupervised Artifact Removal:
Introduced OT-CycleGAN frameworks to remove ultrasound artifacts without matched reference data, offering performance comparable to supervised methods.
Phase Aberration Robust Beamformer for Planewave US Using Self-Supervised Learning
Designed a self-supervised learning approach to robustly generate a high-quality image from various phase aberrated images by modeling the variation in the speed of sound as stochastic.
AI in Bioinformatics and Computational Biology
Protein Sequence Classification
Designed ML-based systems for predicting multiple protein types (e.g., antifreeze proteins, extracellular matrix proteins). These are among the most accurate predictors for analyzing and classifying proteins from their sequence-derived features.
AI in Communication and Adaptive Signal Processing
Adaptive Signal Processing & Learning Algorithms
Explored advanced variants of stochastic gradient descent to speed up convergence and enhance steady-state performance in channel equalization, system identification, and neural network training.
Investigated fractional-gradient approaches, shedding light on their theoretical underpinnings and practical limits.
AI in Surface-Enhanced Raman Spectroscopy
Surface-Enhanced Raman Spectroscopy (SERS)
Applied machine learning for reproducible molecule detection in challenging conditions, addressing longstanding concerns about variability in SERS measurements.
AI in Computer Vision & Model Security
Spatio-Temporal Deep Learning for Robust and Efficient Face Presentation Attack Detection.
•Develop a robust lightweight model for face spoofing detection.
•Extract spatio-temporal features to identify attacks.
•Enable high performance on mobile devices.
Current Work & Goals
Unsupervised & Self-Supervised Learning: Building robust models that do not require paired training data, crucial for real-world biomedical applications where clean or labeled data is scarce.
Generative Modeling: Leveraging diffusion models, transformers, and CycleGANs to improve accuracy and interpretability in imaging tasks.
Runtime Adaptivity: Developing switchable and tunable neural networks that adjust to user preferences on the fly, boosting usability in medical and industrial settings.
Extended Applications: Adapting these frameworks to bioinformatics, computer vision, and surface-enhanced Raman spectroscopy to enhance reproducibility and reliability.
Conclusion
By combining domain knowledge, advanced ML/DL models, and rigorous mathematical frameworks, my work aims to deliver accurate, efficient, and explainable solutions to inverse problems in healthcare, biotechnology, and beyond.
Google Sites
Report abuse