11/03/2026
By Yanhua Si
Details
- Date: Tuesday, Nov. 17, 2026
- Time: 10 – 11:30 a.m. EST
- Location:Zoom meeting
Committee members
- Prof. Yu Cao, co-chair: Miner School of Computer & Information Sciences, University of Massachusetts Lowell
- Prof. Benyuan Liu, co-chair: Miner School of Computer & Information Sciences, University of Massachusetts Lowell
- Prof. Hengyong Yu: Department of Electrical and Computer Engineering, University of Massachusetts Lowell
- Prof. Ming Shao: Miner School of Computer & Information Sciences, University of Massachusetts Lowell
Abstract
Esophagogastroduodenoscopy (EGD) is a primary clinical procedure for the detection and diagnosis of gastric diseases. With the rapid development of deep learning, automated analysis of endoscopic images has become an important research direction for improving diagnostic accuracy and reducing physician workload. However, reliable computer-assisted analysis of gastric endoscopy remains challenging because endoscopic images exhibit large visual variability across anatomical regions, ambiguous boundaries between neighboring gastric sites, subtle lesion appearances, and inconsistent examination quality caused by inadequate gastric preparation.
The first preliminary study focuses on anatomical localization, which is essential for structured examination documentation and region-aware image analysis. We propose a hierarchical deep learning framework for fine-grained gastric region classification in endoscopic images. Instead of treating anatomical localization as a flat multi-class classification problem, the framework incorporates anatomical relationships through a multi-stage hierarchy. A soft-routing mechanism preserves uncertainty during multi-level inference, reducing error propagation and improving classification performance for 12-class gastric region recognition tasks.
The second preliminary study investigates generative models for synthetic data augmentation in gastric lesion detection. By generating additional lesion-related training samples, this study addresses limited annotated data and evaluates whether synthetic images improve model robustness when detecting subtle gastric abnormalities.
Despite these advances, a remaining gap is that most existing methods assume the input images are diagnostically informative and of sufficient visual quality. In real clinical examinations, this assumption often does not hold. Mucus, bubbles, residual food, and poor gastric preparation can substantially reduce image quality and may compromise downstream models for anatomical localization and lesion detection. Therefore, without explicitly assessing gastric cleanliness, automated endoscopic analysis systems may produce unreliable results even when the downstream models are well designed.
To address this gap, the proposed work develops an automated gastric cleanliness assessment framework that evaluates examination quality directly from endoscopic images. The framework will provide an objective and quantitative measure of gastric cleanliness, allowing the system to identify whether an image or examination is suitable for reliable computer-assisted analysis. In this way, gastric cleanliness assessment serves as a quality-control module before downstream interpretation.
The proposed framework has two main objectives. First, it enables automatic quality control by identifying endoscopic images that satisfy diagnostic standards. Second, it improves the reliability of downstream tasks, including anatomical localization and lesion detection, by ensuring that analysis is performed on diagnostically adequate images. Together, the studies in this thesis form a structured deep learning framework for gastric endoscopic image understanding, integrating examination quality assessment, anatomical localization, and lesion detection to support more reliable computer-assisted diagnosis in clinical endoscopy.