Computer vision
Endoscopy AI
Building useful medical computer vision models under limited annotation requires treating labeling strategy and temporal context as modeling decisions.
- endoscopy images
- 3K+
- lesion classifier AUROC
- 0.88
- HRME AUC-ROC
- 0.96
plus dozens of videos
95% accuracy
01
Overview
This work spans two related medical-imaging efforts: a lesion-detection pipeline over endoscopy images and video, and an HRME classifier for distinguishing neoplastic from non-neoplastic Barrett’s esophagus tissue.
02
Approach
For endoscopy data, the pipeline uses pseudo-labeling, active learning, and temporal label propagation to make use of more than 3,000 images and dozens of videos while limiting manual annotation demands. The classifier reached approximately 0.88 AUROC as work progressed toward real-time video inference.
The related HRME model combines ConvNeXt with convolutional block attention modules and reached approximately 95% accuracy and 0.96 AUC-ROC on neoplastic-versus-non-neoplastic classification.
03
My contribution
I built the semi-supervised lesion-detection pipeline, including pseudo-labeling, active-learning, and temporal-propagation components. I also led development of the ConvNeXt and CBAM-based HRME classifier.
04
Scope & limitations
These are research and development results, not claims of clinical deployment. Performance estimates are specific to the evaluated datasets; broader validation and real-time system testing remain distinct requirements.
Methods