Medical Imaging & Image Quality Enhancement

Low-Dose CT Denoising & Perceptual Quality Modeling

Representative LDCT Denoising Results

Context. Multi-year research program focused on low-dose CT denoising and perceptual image quality assessment, integrating supervised, unsupervised, and no-reference IQA methodologies.

  • Problem. Restore high-quality CT images from ultra low-dose acquisitions while preserving diagnostic structures and texture fidelity.
  • Approach. Designed a multi-frame deep learning framework, developed unsupervised training pipelines without clean references, and introduced a self-supervised NIQA metric for perceptual CT quality.
  • Role. Led model architecture design, training pipelines, evaluation on clinical datasets, and coordinated collaboration with radiologists and imaging physicists.
  • Methods. PyTorch, self-supervised learning, CT chracteristic design, noise modeling, perceptual metric design.
  • Outcomes. Multiple peer-reviewed publications and a comprehensive systematic review defining perceptual-quality-driven evaluation standards for LDCT denoising.

📄 Selected Publications

  • A systematic review of deep learning-based denoising for low-dose computed tomography from a perceptual quality perspective (2024) – Biomedical Engineering Letters
  • An unsupervised two-step training framework for low-dose computed tomography denoising (2023) – Medical Physics

🔒 Selected Patents

  • A learning and restoration method for reducing image noise using a neural network
  • Method for removing medical image artifacts using artificial intelligence

➡️ Full publications & patents

Physics-Informed Neural Adaptive Optics for OCT 🔬 (Ongoing)

Self-supervised deep learning framework for computational aberration correction in optical coherence tomography—eliminating the need for expensive hardware adaptive optics.

  • Problem. Optical aberrations degrade OCT image quality, limiting achievable resolution. Hardware AO is expensive ($50K–200K+) and assumes phase stability that many systems lack.
  • Solution. Designing a physics-informed neural field approach that jointly learns tissue structure and optical aberrations from a single measurement volume—fully self-supervised, no ground-truth labels required.
  • Core Innovations.
    • Neural implicit representation for complex-valued scattering fields
    • Differentiable OCT forward model with coherent interference physics
    • Joint optimization of structural MLP weights + Zernike-based aberrations
  • Stack. PyTorch, custom OCT physics simulator
  • Current Progress.
    • Neural coordinate representations adapted to interferometric imaging
  • Roadmap.
    • 🩻 Successful phantom/tissue validation with measurable resolution improvement
    • 🔬 Exploration of depth-dependent aberration modeling
    • 🧠 Real-time inference for clinical deployment
    • 📊 Extensions to PS-OCT and OCTA
    • 📄 Publication target: biomedical optics journal (2026)

The model jointly estimates the underlying 3D scattering structure and depth-dependent optical aberrations using a differentiable forward model grounded in OCT interferometry. This enables coherent reconstruction without hardware adaptive optics.

[Code, Preprint — Available upon publication]

Psychology RAG System 🧠 (Ongoing · Independent Project)

A self-directed, independent research project exploring how large language models and retrieval-augmented generation (RAG) systems can function as AI-assisted reasoning tools for synthesizing fragmented interdisciplinary knowledge into coherent theoretical frameworks.

  • Problem. Complex psychological phenomena (e.g., trauma responses, behavioral escalation loops, emotional incongruence) are distributed across hundreds of papers, books, and clinical frameworks. Conventional keyword-based search and linear reading fail to expose latent conceptual links and temporal dependencies, obstructing theory formation.

  • Solution. Designed and implemented a 3-tier Hierarchical RAG architecture operating as an expert reasoning partner (human-in-the-loop), enabling structured contextual reframing and cross-domain synthesis across psychology, neuroscience, and computational modeling.

  • Technical Stack. Python, FastAPI, React/TypeScript, ChromaDB, OpenAI Embeddings, GPT-4 / Claude

  • Research Outcomes.

    • Generated novel conceptual frameworks using the system as an AI-augmented research instrument, including:

      • Shame-Driven Dopamine Loop
      • Five-stage behavioral escalation pathway
    • White paper preprint submitted to arXiv (currently under moderation)

    • Demonstrates practical capabilities in:

      • Hierarchical RAG design
      • Interdisciplinary knowledge synthesis
    • Initial system scale: 9,610 indexed chunks from 330+ documents, achieving sub-2s retrieval latency

  • Roadmap.

    • 🧠 Personalized psychological analysis and reasoning engine
    • 🗂️ Hierarchical knowledge graph with semantic cross-referencing
    • 🎯 Adaptive retrieval strategies (dynamic weighting by context and document type)
    • 🪄 LLM reasoning orchestration
    • 🔀 Query orchestration pipeline: decomposition → parallel retrieval → aggregation

Code: Private repository (under active development) · White Paper: Preprint available