Sangwoo Ha

AI Researcher

I am an AI researcher working on medical image analysis and on-device inference. Previously, I developed cervical cytology models at Noul, chest X-ray models at DeepNoid, and industrial video analytics at HumanICT. I also develop open-source AI tools.

Medical imagingComputer visionOn-device AI

Research & development ↓

Research & Development

Cervical cells from Noul’s official miLab CER product materials

Noul · Product image ↗

NOUL · 2023–2026 In product

miLab™ CER · Cervical Cytology

I was the sole developer of the YOLO-based cell detector used in miLab CER. I supplemented the cell annotations for training and compressed the model to run on the instrument’s Jetson Nano and TX2 hardware.

Detection / Pseudo-labeling / Quantization / Jetson

Contributions

Used CellSAM and Cellpose with pseudo-labeling and teacher–student training to supplement cell annotations.
Quantized the model and optimized inference for Jetson Nano and TX2.

I also researched Diffusion, DiT, and StyleGAN models for generating training samples, and designed DINOv3-based models and report-generation architectures using MedGemma and LLMs.

Representative chest X-ray from DEEPNOID’s official DEEP:CHEST product page

DEEPNOID · Product image ↗

DEEP:NOID · 2020–2022 Product & clinical study

DEEP:CHEST · Chest X-ray Analysis

I designed models to detect pneumothorax and consolidation and locate lesions in chest X-rays. I led the three-person DEEP:CHEST team and took part in the clinical trial, regulatory approval, and product development of DC-XR-03.

Chest X-ray / Localization / Clinical validation

Contributions

Developed a TransUNet-based semi-supervised approach for chest X-rays with incomplete mask annotations, modifying the loss and layers to learn from both pixel-level and classification labels.
Published the clinical validation study in npj Digital Medicine as a co-first author.
Developed models for the follow-up DC-XR-04 to detect and localize multiple thoracic conditions, including pulmonary nodules and pleural effusion.

Alphai PlantVision · Industrial Video Analytics

I was responsible for AI model development for construction-site CCTV, detecting fire, smoke, and missing safety helmets. I reduced the computational demands of inference with TensorRT INT8 and worked with Hyundai Engineering & Construction on a PoC.

Object detection / TensorRT / Video analytics

Contributions

Developed detectors for fire, smoke, missing safety helmets, construction vehicles, and general vehicles.
Compressed models and converted them to TensorRT INT8 for real-time inference with limited compute resources.
Used pixel differences between frames in post-processing to improve detection.

Publications

npj Digital Medicine · 2022 Co-first author

Localization-adjusted diagnostic performance and assistance effect of a computer-aided detection system for pneumothorax and consolidation

Sun Yeop Lee*, Sangwoo Ha*, et al. · * Equal contribution

A clinical study evaluating lesion localization and the effect of AI assistance on physicians’ interpretation of chest X-rays for pneumothorax and consolidation.

Read the paper ↗