By Dave DeFusco
Colorectal cancer remains one of the deadliest forms of cancer worldwide, claiming hundreds of thousands of lives each year. The key to reducing its toll is early detection—specifically, the identification and removal of precancerous polyps during routine colonoscopies. However, polyp detection is an arduous task. Their varying size, shape, color and texture, along with visual obstructions like reflections and bowel contents, make them easy to miss.
To address this challenge, Katz School researchers have developed PolypSEAG-Net—a novel deep learning model that enhances polyp segmentation in colonoscopy images. Their work, recently presented at the ACM/IEEE International Conference on Connected Health, will contribute to the advancement of medical image analysis and colorectal cancer detection.
“Despite advancements in artificial intelligence and deep learning, automatic polyp detection remains difficult,” said Dengyi Liu, lead author of the study, a Katz School Ph.D. student in mathematics and a 2024 graduate of the M.S. in Data Analytics and Visualization. “Traditional computer-aided detection methods rely on handcrafted features, which often fail to generalize across different datasets and clinical conditions.”
Deep learning models, particularly convolutional neural networks (CNNs), have significantly improved medical image segmentation, yet many existing architectures struggle with false positives, poor boundary delineation and difficulty generalizing across diverse polyp appearances. PolypSEAG-Net builds on the strengths of established deep learning architectures while integrating two powerful techniques—attention gates (AGs) and squeeze-and-excitation (SE) blocks—to refine feature extraction and improve accuracy.
The model’s key contributions include:
- Enhanced Feature Extraction: PolypSEAG-Net employs advanced convolutional layers to capture intricate details of polyps, reducing false positives and improving segmentation accuracy.
- Attention Gates (AGs): These mechanisms enable the model to focus on the most relevant regions of an image, filtering out irrelevant background noise and distractions.
- Squeeze-and-Excitation (SE) Blocks: SE blocks improve channel-wise feature selection, ensuring that the most critical visual features are emphasized while suppressing less useful information.
- Robust Data Augmentation: To increase the model’s generalization capabilities, the researchers implemented extensive data augmentation techniques, making PolypSEAG-Net more adaptable to variations in polyp size, texture and imaging conditions.
“To evaluate its performance, we tested PolypSEAG-Net on three publicly available datasets: ClinicDB, PolypGen and Kvasir-SEG,” said Dr. Ming Ma, senior author and an assistant professor in the Graduate Department of Computer Science and Engineering. “The results were striking. The model outperformed state-of-the-art segmentation models, including TransUNet and UniverSeg, across most key evaluation metrics such as Dice similarity coefficient and Intersection-over-Union.”
These findings suggest that PolypSEAG-Net is not only an improvement over classical models but a potential new standard for automated polyp detection in colonoscopy. The implications of the research extend beyond improved segmentation metrics. In clinical practice, higher accuracy in polyp detection can directly lead to earlier interventions, reducing the risk of polyps developing into cancerous tumors. A model like PolypSEAG-Net could be integrated into existing colonoscopy systems, assisting gastroenterologists in real-time by highlighting suspicious polyps, thereby reducing the chance of human error.
“The model’s efficiency and adaptability make it a strong candidate for low-resource settings, where access to highly trained specialists may be limited,” said Dr. Hua Fang, a co-author of the study and professor of computer and information science at the University of Massachusetts (UMass) Dartmouth and UMass Chan Medical School. “By providing AI-assisted diagnostics, PolypSEAG-Net has the potential to democratize access to life-saving colorectal cancer screening.”
The researchers see room for further improvement. Future iterations of PolypSEAG-Net may incorporate transformer-based architectures, meta-learning techniques for adapting to unseen datasets and self-supervised learning to reduce reliance on labeled training data. Another avenue of research includes integrating diffusion models to enhance feature extraction in complex polyp segmentation tasks.
“By leveraging advanced deep learning techniques, our research is setting a new benchmark in medical image analysis,” said Dr. Honggang Wang, a co-author of the study and chair of the Graduate Department of Computer Science and Engineering. “If implemented clinically, PolypSEAG-Net could save countless lives by making polyp detection faster, more accurate and more reliable than ever before.”
