Smarter AI for early cancer detection in Barrett鈥檚 Neoplasia
Tim Boers defended his PhD thesis at the Department of Electrical Engineering on July 1st.
Barrett鈥檚 Esophagus (BE) is a precancerous condition in which the normal lining of the lower esophagus is replaced by specialized intestinal tissue, increasing the risk of developing esophageal adenocarcinoma鈥攁 cancer with poor prognosis when detected late. Early detection of neoplastic lesions in BE is critical for enabling minimally invasive treatment and improving patient outcomes. This research of Tim Boers contributes to that goal by developing a clinically validated, efficient, and real-time computer-aided detection (CADe) system that can be embedded into existing endoscopic equipment. By combining low-complexity algorithms, robust training techniques, and stable visual feedback, the system delivers expert-level diagnostic support with minimal computational demands鈥攎arking a significant step toward routine clinical adoption of AI-assisted endoscopy.
While deep learning has led to remarkable advancements in computer-aided detection (CADe), many of these systems remain too complex for integration into real-world clinical environments. Most high-performance models demand substantial computational resources and memory, making them difficult to embed into the compact, real-time systems used during endoscopic procedures.
Designing an embedded, efficient CADe system
This research of presents a major step forward in the design of AI systems for medical imaging by focusing on efficiency, robustness, and seamless integration. A key innovation is the development of a compact, quantized algorithm that drastically reduces system complexity. Thanks to its low computational demands, the CADe system can run in real time on existing embedded programmable chips, making it suitable for integration into current endoscopic video equipment.
Clinically validated, expert-level performance
To ensure clinical relevance, the developed system underwent extensive validation in multiple clinical studies. These studies demonstrated that the CADe system not only performs with high accuracy across a wide range of subtle lesion types, but also significantly enhances the diagnostic performance of general endoscopists鈥攅levating their detection capabilities to near-expert level.
Boosting robustness with innovative AI techniques
Robustness was addressed through several technical strategies. One important contribution is the successful use of self-supervised learning for data augmentation, which improves performance without requiring additional labeled data. Moreover, the research demonstrated the value of gastrointestinal endoscopic data for pretraining CADe systems, providing a strong foundation for accurate lesion detection.
Temporal stability through recurrent models
A final contribution lies in the improvement of temporal consistency of detection results. By incorporating recurrent neural network (RNN) models, the system generates smoother, more continuous visual feedback during endoscopic procedures, helping clinicians maintain focus and interpret results more effectively in real time.
Towards clinical integration of AI in endoscopy
This research delivers important contributions to the development of robust, efficient, and embedded AI systems for the detection of neoplasia in Barrett鈥檚 Esophagus. The CADe system combines high performance with low complexity, has been clinically validated, and is now approved for embedded system integration by a leading manufacturer of endoscopic equipment. These results represent a significant step toward widespread clinical acceptance and adoption of AI-supported diagnostics in gastroenterology.
Title of PhD thesis: . Supervisors: Prof. Peter de With (果冻传媒), Prof. Jacques Bergman (Amsterdam UMC), and Dr. Fons van der Sommen (果冻传媒).