Yi Qin (Eason)

I'm now currently a Ph.D. candidate supervised by Prof. Xiaomeng Li at ECE, the Hong Kong University of Science and Technology. I also had the opportunity to work with Prof. Hao Wang and Prof. Lu Mi. Before joining HKUST, I obtained BEng at the South China University of Technology majoring in Automation Science and Engineering.

Email  /  CV  /  Scholar  /  Twitter  /  Linkedin

Research

My research focuses on leveraging generative models to enhance the quality and trustworthiness of digital healthcare, with a particular emphasis on cardiology and echocardiography. Specific interested topics include:

• Multi-Modal Large Language Model / Foundation Model for Disease Diagnosis/Prognosis
• Trustworthy/Explainable DL (Energy-based Concept-based models)
• Diffusion-based Generative Model

Cardiology

EchoChat integrates multimodal echocardiography with clinical knowledge and visual evidence to support cardiac condition recognition, measurement, report generation, and conversational interpretation.

CardiacPULSE detects end-diastolic and end-systolic frames through self-supervised learning of cardiac motion periodicity. Fourier-guided spatial priors and a lightweight network enable efficient phase detection across echocardiographic views.

CardiacMind combines a Cardiac Reasoning Template with reinforcement learning rewards for procedural detail, cross-view evidence integration, and visual grounding, enabling interpretable reasoning for complex echocardiographic diagnosis.

We propose the first Multimodal Medical Reasoning-with-Retrieval framework, which actively retrieves external knowledge by querying observed symptoms or domain-specific medical concepts during reasoning.

Echo-Path generates echocardiogram videos conditioned on specific cardiac pathologies. Its pathology-conditioning mechanism allows the model to learn and control disease-specific structural and motion patterns in the heart.

We propose Multi-Agent Collaborative Expertise Extractor, a multi-agent system that builds EchoCardiography Expertise Database, the richest cardiac knowledge base from diverse sources. We also introduce Echocardiography Expertise-enhanced Visual Instruction Tuning, a lightweight tuning method that efficiently injects this expertise into models by training less than 1% of parameters.

EchoViewCLIP addresses these issues using a large dataset with 38 standard views and OOD samples. It introduces a Temporal-informed Multi-Instance Learning (TML) module for capturing key frames and a Negation Semantic-Enhanced (NSE) detector for OOD rejection. A quality assessment branch boosts reliability. The model achieves 96.1% accuracy, advancing fine-grained view recognition and robust OOD handling in echocardiography.

To fully exploit the diffusion model's ability to guide the registration task, we present two modules in FSDiffReg: Feature-wise Diffusion-Guided Module (FDG) and Score-wise Diffusion-Guided Module (SDG).

Interpretation

CUDA improves the robustness of Concept Bottleneck Models under domain shifts by aligning concept representations with adversarial training, allowing flexible differences, and enabling concept inference without labels. It boosts interpretability and outperforms state-of-the-art CBM and domain adaptation methods.

We propose Energy-Based Conceptual Diffusion Models (ECDMs), a framework that unifies the concept-based generation, conditional interpretation, concept debugging, intervention, and imputation under the joint energy-based formulation.

We introduce Energy-Based Concept Bottleneck Models (ECBM) as a unified framework for concept-based prediction, concept correction, and fine-grained interpretations based on conditional probabilities.

Digital Healthcare

TriALS benchmarks liver lesion segmentation when contrast-enhanced CT is unavailable. The multicenter challenge provides four-phase CT data and evaluates how methods transfer to non-contrast scans and independent external cohorts.

We introduce CvG-Diff, a fast, high-quality sparse-view CT reconstruction method that models artifacts deterministically and introduces two novel techniques—EPCT and SPDPS—to reduce errors and improve efficiency.

XMedGPT combines textual and visual explanations with uncertainty assessment to support trustworthy medical decision-making. It grounds anatomical references in images and supports diagnosis, report generation, and prognostic modeling across diverse imaging modalities.

SyRe is a text-prompted segmentation foundation model trained on 20 million image-mask-description triplets across 9 modalities and 229 tasks. It supports flexible segmentation requests, invalid-prompt rejection, and clinician-guided refinement, with validation across 28 unseen external datasets.

The ACR-NCI-NVIDIA challenge evaluates federated learning for breast density classification across mammography datasets from multiple sites, comparing generalization while keeping training data at their original institutions.

Projects and Patents

Honors, Awards, and Services

This website is adapted from Jonathan Barron's source code.