Yong-Gi Hong

Owned the full AIRECIST pipeline for automated lung-cancer RECIST response assessment.

• Reframed a reproducibility problem (inter-reader RECIST agreement below 70%) into four stages: lesion segmentation → pre/post registration → tumor tracking → automated RECIST decision
• Combined nnUNet-based segmentation with ROI preprocessing (DICOM→NIfTI, lung-window HU clipping); implemented VoxelMorph-based registration and a custom tumor-tracking algorithm
• Result: lesion Dice 82%, RECIST auto-decision 91% accuracy, clinically validated at Asan Medical Center (Dept. of Radiology)
• Ontol for Clinics (AI health-checkup reports): contributed the algorithm that drafts physician findings from checkup values in the EMR database
• TIPS project (breast-cancer report analysis & personalized MRI 3D visualization): NLP report summarization, plain-language medical term rewriting, and tumor / lymph-node / body segmentation on breast MRI
• Contributed to a MOTIE national R&D project (cardiovascular biomechanics digital simulator)

Developed and productized commercial segmentation, GAN, and registration models.

• DeepCATCH (commercial): whole-body CT multi-organ segmentation with 30+ classes using 2D/3D nnUNet, SwinUNETR, U-Net — consistent quality across large multi-center data
• TiSepX (commercial): X-ray tissue separation GAN (bone/lung/pulmonary) — benchmarked CycleGAN, DP-GAN, BMD-GAN, OASIS; model compression for fracture detection and DEXA BMD use
• Synthesized X-ray images from CT volumes (DRR-based) to expand the training set
• Lung / Multi-phase Registration (registered patent, co-inventor): Lung Seg → Affine → DL registration (VoxelMorph-based cascade), patch training with sliding-window inference
• Result: quantified prone-posture lobar volume gain of +90.2 mL (LLL) / +52.5 mL (RLL), established an LLL volume–FVC correlation

Multi-modality medical imaging research.

• EEG breathing-pattern classification with LDA/Random Forest → 1st-author SCI paper (Brain Sciences, 2021)
• AI-Dentis (malocclusion & cephalogram): Inception-ResNet classifier with Grad-CAM → novel first-molar relative-distance classification (master's thesis, patent, web commercialization); 46-landmark cephalogram detection and a labeling tool
• PET reconstruction (SITL): physics-informed lightweight model — 247K parameters (vs. DeepPET's 63.6M → 257× lighter), SSIM 0.983 / PSNR 36.6 dB, validated with a self-fabricated PET phantom