AI Researcher · Seoul
Computational Pathology / Pathology Foundation Models / Multi-Omics
Five years building deep learning for medical imaging and drug discovery: from whole-slide-image pipelines and MFDS-track validation to pathology foundation models, now integrating proteomics and multi-omics.
I build deep learning systems for computational pathology and drug discovery. First-author work on H&E-based breast-cancer recurrence in Scientific Reports, a co-first review in Prostate International, and a contribution to large-scale pathology foundation models (G2L, AAAI 2026 workshop, oral).
At OMIXAI I'm extending histopathology AI into proteomics and multi-omics, including a Korea–US–Japan precision-oncology collaboration on Pan-Sarcoma proteogenomics. A recurring thread is predicting molecular state from H&E alone (gene expression, spatial proteomics, receptor status) and testing whether those predictions can actually stand in for a molecular assay. Earlier, at Deep Bio, I led WSI pipeline development and pre-deployment model validation for an MFDS-track product.
Scientific Reports · 2025 · First author
Multi-center retrospective study (n=125; National Cancer Center + Korea University Guro) showing histopathology-only deep learning can approximate genomic-assay-style risk stratification.
AAAI 2026 Workshop (W3PHIAI) · Oral
Distilling giga-scale pathology foundation models into compact, cancer-specific encoders via knowledge distillation, preserving downstream performance at lower compute.
Medical Image Analysis · 2026
The official challenge report (Medical Image Analysis, vol. 114, 104234) for kidney glomeruli segmentation, 2nd place in the whole-slide track, tackling generalization from patch-level training to slide-level inference.
Prostate International · 2025 · Co-first
A review of current trends and future directions for AI in urological cancer pathology, spanning clinical, technical, and regulatory perspectives.
Bioengineering · 2024
Ingests multiple resolutions simultaneously and abstains on unreliable regions via a confidence-based reject loss, built for noisy histopathology labels.
Preprint · 2026
Using H&E-based AI risk scores to guide spatial proteomics across 156 triple-negative breast cancer patients. High-risk niches showed proliferative programs and low-risk niches immune activation, yielding a 13-protein signature that improved prognosis when combined with the image-based risk score.
Preprint · 2026 · Co-first
A linear framework that groups WSI patches into morphological microstates, discovers spatial macrostates from their adjacency, and maps them through low-rank molecular programs to gene expression. Outperforms 14 baselines across three TCGA cancers, stays strong with less training data, and its linear form attributes each prediction to global and region-specific programs.
Preprint · 2026
Efficient AI-driven analysis across multiple tissue sections to predict biochemical recurrence, extending single-slide inference to full multi-section specimens.
IEEE Access · 2021
Region-annotation-level pixel-wise contrastive loss to stabilize segmentation under scarce and unreliable medical-image labels; the line of work behind my M.S. thesis.
Seoul · Nov 17–20, 2026. High accuracy does not make an H&E model a substitute for a molecular test. We define substitution cost by translating each misclassification into a deviation from a prespecified treatment routing, and apply one pre-registered protocol (UNI + CLAM-SB, site-disjoint hold-outs, label-shuffle controls) across five cancers. Of ~15 endpoints only HPV status in head and neck met the confirmation criterion; most were reported as undecided rather than negative. With Pseudo Lab.
Seoul · Nov 17–20, 2026. Are the per-gene parameters that multiome RNA-velocity methods emit (transcription rate α, degradation rate γ, chromatin-to-transcription lag) actually reliable? Five velocity arms tested on four axes: cross-method reproducibility, an ATAC-shuffle control, replication in five external multiomes, and anchoring to measured synthesis and degradation rates. Only α reproduced across methods; γ and the lag need orthogonal validation.
AACR Annual Meeting 2023 · Poster · First author
Abstract #5404 · April 2, 2023. ER, PR and HER2 status from H&E WSIs alone, without IHC. TCGA-BRCA (728 cases), a multi-task model with a shared morphology encoder and per-receptor heads, plus confidence-based patch selection. With Korea University Guro Hospital.
AACR Annual Meeting 2022 · Poster · First author
New Orleans · April 12, 2022. Can H&E WSI analysis approximate the 21-gene recurrence score? 125 hormone-positive, node-negative, HER2-negative cases; patch-level CNN with majority voting into low / intermediate / high risk, nested cross-validation. No low-risk case was called high-risk. The work that grew into the Scientific Reports (2025) paper.
AACR Annual Meeting 2022 · Poster · Co-author
New Orleans · April 2022. Survival risk from pancreatic adenocarcinoma H&E, with a general adenocarcinoma feature extractor (GAFE) to test transfer to other organs. TCGA pancreas (179 cases), 5-fold CV; transfer evaluated on rectum (READ) and breast (BRCA).
USCAP Annual Meeting 2022 · Poster · Co-author
March 18, 2022. Breast-cancer death risk from H&E WSIs using features from a pre-trained prostate diagnosis model instead of a general-image backbone. TCGA-BRCA (980 WSIs), attention-guided multiple-instance learning, 5-fold CV; Kaplan-Meier separation of high- vs low-risk groups.
USCAP Annual Meeting 2022 · Poster · First author
March 17, 2022. Consistent histology grading (tubule formation, nuclear grade, mitoses) to reduce inter-observer variability. 125 H&E WSIs graded and annotated by a pathologist; tumor patches classified into grade 1 / 2 / 3, with class-activation maps for inspection. With Korea University Medicine.
KIIE Fall Conference 2020 · Oral · First author
With Prof. Sangheum Hwang, Dept. of Data Science, SeoulTech. A feature-wise contrastive loss on the encoder bottleneck pulls same-class embeddings together and pushes different classes apart. Gains for both U-Net and UNet++ on CT lung (Kaggle) and liver (LiTS2017) segmentation, largest on distance-based boundary metrics. The work behind the M.S. thesis and the IEEE Access (2021) paper.
Self-supervised proteomic representation learning with LoRA / PEFT fine-tuning; also cell-line IC50 prediction. Validated under a leave-drug-out protocol on held-out compounds.
Canine oncology drug-recommendation algorithm toward a clinical decision support system, plus co-leading the Virtual Cell Challenge, predicting CRISPR-knockdown response in pluripotent stem cells.
NVIDIA Korea Agentic AI Hackathon 2026 entry (team of 5) linking pre-market target binding (FlyDiscovery) and post-market adverse-event surveillance (FlyVigilance) in one agent. I built FlyDiscovery, which scores candidates from structure prediction to affinity with BioNeMo NIM (OpenFold3 · DiffDock · Boltz-2), and its discovery-workbench UI.
Predicting breast-cancer prognosis from H&E WSIs only, with no other clinical inputs. Patch-wise classification with confidence-aware selection and majority voting into slide-level GHI-RS risk groups.
ADMET prediction and drug–target interaction modeling for early drug-property profiling, plus an active-learning framework that defines out-of-distribution pathology samples to make labeling more efficient. Internal R&D, no public results.
2D and 3D segmentation of hemorrhage regions plus slice-level classification. Validated domain generalization on external public data and used class-activation maps to audit classification reliability (U-Net vs DeepLab).
Defect detection on a manufacturing line: 969 images across 5 classes, under a hard 0.3s/image budget. Used machine vision (not DL) to localize the screw and strip background before classification, which lifted accuracy.
Transaction-price modeling for Seoul/Busan from 1.6M records. Re-matched addresses by postal code to fix mislabeled districts, and enriched with public-API features (nearby subway, schools) missing from the raw data.
Daily Top-100 charts from 6 streaming platforms (Melon, Genie, VIBE, YouTube Music, Apple Music, Bugs) crawled at a fixed time, merged on track + artist, and loaded into a relational DB. Schema designed for chart-in/out sparsity (nullable rank columns) and multilingual titles (nvarchar).
U-Net segmentation on PASCAL VOC (20 classes, 1,928 images, 4,203 objects) with adversarial training, label smoothing and cut-out applied alone and in combination, comparing the robustness–accuracy trade-off; the starting point of the robustness-first view carried into medical AI.
KPI Challenge 2024 · Whole-Slide Track
Glomeruli segmentation · held at MICCAI 2024 · published in Med. Image Anal. 2026
Foundation models, multi-omics, or translational digital pathology: I'd like to hear about it.