Index

Development of a Local LLM Agent System for Clinical Expert Support and Automation in MRI Planning for Radiation Therapy

RAG-Enhanced Low-Cost Vision-Language Models for Diabetic Retinopathy Classification and Automated Reporting

Diabetic Retinopathy (DR) affects over 160 million people globally, projected to reach 180 million by 2030, despite 90% of related blindness being preventable through early detection [1]. Current AI models achieve strong classification performance but lack interpretable clinical reports, limiting their adoption in low-resource settings. Although Vision-Language Models (VLMs) offer unified diagnosis and report generation, fundus captioning significantly underperforms compared with other imaging modalities [2,3], and state-of-the-art VLMs remain computationally expensive. Although Retrieval-Augmented Generation (RAG) has improved medical imaging accuracy [4], no prior study has integrated DR severity grading, lesion-aware reporting, and evidence retrieval within a low-cost, clinically deployable VLM.

Generalization of Self-Supervised Vision Models in Image Retrieval

This project investigates the generalization capabilities of prominent self-supervised vision models, such as DINOv2, CLIP, and MoCo, when applied to image retrieval tasks across diverse visual domains.

Overview

Self-supervised learning (SSL) models are increasingly important for creating robust visual features. However, their performance often degrades when transferring from standard natural image datasets (like ImageNet) to more specialized domains.

Our study systematically benchmarks these models on three distinct domains:

1. Natural Images (Baseline)

2. Scene-Centric Images (e.g., Places365)

3. Artistic Images (e.g., ArtPlaces)

We compare the models’ out-of-the-box generalization and analyze the impact of finetuning on their domain adaptation, providing crucial insights into the stability and robustness of the learned representations for practical retrieval applications.

Thesis_Proposal__Workflow_2