JIN HYUN KIM

Medical AI Projects

This page summarizes ongoing and past projects in medical AI, with a particular focus on ophthalmology, retinal imaging, and trustworthy clinical decision support.

Key Themes

• Domain-specialized LLMs for ophthalmology (e.g., Ophtimus-V2-Tx)
• Noise-robust medical image analysis and quantification
• Reliable mapping from model outputs to clinical coding systems
• Evaluation frameworks for safety, robustness, and explainability

Selected Projects

• Ophtimus-V2-Tx: An 8B-parameter ophthalmology LLM trained on case reports and evaluated with CliBench-based coding.
• ERM Quantification: Low-cost and fast SD-OCT based epiretinal membrane detection and thickness quantification.
• Trustworthy Clinical Support: Methods to validate and monitor LLM predictions before deployment in real clinical workflows.

Ophtimus: Ophthalmology-specific LLM

Python · PyTorch · Transformers · LangChain · Streamlit · FastAPI

🤗 Models and Datasets   |   📕 AAAI 2025 Workshop Paper

Introduction

Ophtimus is an open-source large language model (LLM) specialized in ophthalmology, built with 8 billion parameters based on the LLaMA architecture. It is trained on carefully curated ophthalmology-specific data, including medical papers, textbooks, and research reports. Through filtering, summarization, and preprocessing, only the most relevant and high-quality information was retained.

Designed to be both lightweight and high-performing, Ophtimus is suitable for real-world applications such as clinical decision support, medical education, and patient communication. The model and its training pipeline are fully open-sourced, providing a practical reference for developing similar domain-specific LLMs in other areas of medicine.

Related GitHub Repositories
• Ophtimus-Ophthalmology-LLM
• SD-OCT-ERM-Quantification

---

Dataset Details

All datasets used for Ophtimus were either newly constructed or adapted for this project. Pre-training datasets were curated from open-source ophthalmology materials, while instruction-tuning and evaluation datasets were obtained by extracting only ophthalmology-relevant samples from broader medical corpora. All data underwent preprocessing steps, including deduplication, English-only filtering, and removal of any personally identifiable information (PII).

---

Model Details

The pre-training and instruction-tuning columns below refer to the training conducted in this project. The base models had already undergone their own pre-training and/or fine-tuning, and Ophtimus was built using transfer learning on top of these models.

---

Performance

Multi-Choice QA: Ophtimus-Eval, MedMCQA, PubMedQA (ophthalmology-subset)
Essay QA: MedQuAD, Medical Flashcards, Medical Wikidoc (ophthalmology-filtered)

Ophtimus-Eval is a proprietary dataset collected from a medical platform. The other datasets are established medical benchmarks, from which only ophthalmology-related QA pairs were extracted for evaluation.

SD-OCT-based Epiretinal Membrane Diagnostic Assistant System

Python · PyTorch · OpenCV · YOLO · Pillow

Introduction

This project presents a low-cost and efficient method for detecting and quantifying Epiretinal Membranes (ERM) using Spectral-Domain OCT (SD-OCT). Using deep learning techniques—particularly YOLO object detection—we generate en face ERM Projection Images from B-scan data, enabling intuitive visualization and accurate measurement of ERM lesions.

The proposed approach also quantifies the association between ERM severity and retinal thickness, contributing toward enhanced clinical decision-making. This system aims to reduce the diagnostic gap between SD-OCT and Swept-Source OCT (SS-OCT) while maintaining accessibility and diagnostic performance.

GitHub repository: github.com/jinkimh/SD-OCT-ERM-Quantification

YOLO Model Evaluation

We evaluated YOLOv5, YOLOv8, and YOLOv11 models for ERM detection. Each model was trained with two dataset scales (Full: 2200 images, Half: 1100 images) and tested on 650 expert-labeled OCT B-scans.