SEUNGMIN LEE

• Deployed models via Flask REST API for DR/DKD/Stroke risk modules (DeepDR Plus / DeepDKD / DeepRETStroke): request validation, preprocessing/postprocessing, GPU inference pipeline.
• Wrote production inference code (batching, timeouts, error handling, JSON schema) and packaged endpoints for internal consumers.
• Designed multiple endpoints (e.g., /predict_dr, /predict_dkd, /predict_stroke, /triage) with consistent response spec (probabilities, risk tiers, thresholds).
• Built service glue: auth tokens, basic logging, configurable thresholds; prepared API docs (OpenAPI/Swagger) for front-end/ops handoff.

• -Vision–Language Medical AI for Ophthalmic Screening and Service Deployment
• Built a RETFound + ViT–based multi-label classifier for 13 ophthalmic diseases (e.g., DR, DME, glaucoma, AMD) based on visual and linguistic analyses.
• Merged heterogeneous datasets (APTOS2019, IDRiD, RFMiD, BRSET); rebuilt labels/taxonomy and standardized preprocessing.
• Addressed class imbalance with Focal Loss, class weights/pos_weight, and label smoothing; performed per-class threshold tuning for better sensitivity–specificity trade-offs. Replaced [CLS] pooling → Global Average Pooling, improving representation quality and achieving AUC = 0.9569 on the held-out set.
• Deployed a Flask REST API that accepts image uploads and returns predictions; built the service at (www.care-o.chat/login) for end-to-end use.
• Built an LLM RAG corpus from ophthalmology textbooks/guidelines (cleaning→chunking→embedding index) for

— Data Validation & Model Verification (in collaboration with Researcher Sujeong Song)

• Collaborated with Kangbuk Samsung Hospital to organize ophthalmic imaging data and verify labels for quality and consistency.
• Worked with Professor Sujeong Song to design and execute model performance evaluation and validation procedures; ran validation experiments and documented findings.

• -Segmentation-based Brain MRI Alignment and Fusion with Deep Learning
• Built an end-to-end pipeline for brain MRI segmentation, registration, and fusion, enabling lesion-to-baseline comparison to support diagnosis, surgical planning, and treatment assessment.
• Implemented FCN for high-precision anatomical segmentation and combined U-Net–based non-rigid registration with feature-point matching for flexible alignment under varying scan conditions.

• The research about current challenges in constructing a brain using neural networks and about some deep neural network training algorithm such as BP, FA, DFA (python)

• -Multi Radar Data Fusion Project -
• 《Using Micro-Doppler Figures to Identify the Characteristics of Humans and Pets in Indoor Spaces》
• Mainly responsible for writing MATLAB scripts to constructing the CNN model and taking performing the data refinement process from data such as cutting long Micro-Doppler figure data into short Micro-Doppler figure data and deleting useless data, fitting to the machine learning input port.
• Coordinate and lead the project progress, communicate with the Intel teachers on project progress and problem-solving as a team representative.
• Our group won the top three scores in the school-enterprise innovation of Shanghai Jiao Tong University, and was recommended by the school to participate in the Chinese national innovation project.