Optimizing AI-systems for the detection and diagnosis of gastrointestinal cancers
Koen Kusters defended his PhD thesis with the distinction cum laude at the Department of Electrical Engineering on November 11th.
Gastrointestinal (GI) cancers remain among the deadliest cancers worldwide, with esophageal adenocarcinoma and colorectal cancer being of particular concern. These cancers have low five-year survival rates due to late-stage diagnosis, which limits treatment options. However, many GI cancers begin as identifiable precancerous conditions, such as Barrett’s esophagus (BE) or colorectal polyps. Detecting these conditions early enables minimally invasive treatments and improves patient outcomes. In his PhD research, Koen Kusters explores how advanced AI can be support the early detection of GI cancers.
Deep learning has shown great promise in Computer-Aided Detection and Diagnosis (CADe/CADx) systems for identifying early signs of cancer during endoscopic procedures. However, widespread clinical adoption remains limited. Key challenges include a lack of robustness to domain shifts, caused by variability in endoscopic imaging conditions, and the technical and operational demands of integrating AI into real-time clinical workflows. Koen Kusters’ research addresses these challenges, aiming to develop CADe/x systems that are not only accurate, but also robust, efficient, and ready for real-world deployment.
Benchmark study
Kusters began his research by evaluating various foundational deep learning-based architectural families in a comprehensive benchmark study across diverse endoscopic applications. The goal was to assess how well each architecture handles the variability found in endoscopic imaging. This analysis provided valuable insights into which models are best suited for building reliable CADe/CADx systems.
Imaging conditions
Next, Kusters investigated how domain shifts affect CADe/x system performance. These domain shifts mean that when a model trained on one set of images performs poorly on another due to differences in imaging conditions. Modern endoscopy devices offer various image enhancement settings, which can drastically affect image appearance. The study found that these variations can significantly impact model performance. To address this, Kusters introduced a data augmentation strategy that simulates a wide range of enhancement settings during training. This approach made the models more resilient to real-world variability, improving their consistency across different imaging setups.
A next-generation CADe system
Based on the previous insights, Kusters designed, developed and evaluated a next-generation CADe system for BE neoplasia detection. The system was designed with robustness in mind, using self-supervised domain-specific pretraining and a hybrid Convolutional Neural Network-Transformer architecture in combination with a diverse training dataset. The optimized model significantly improves performance and robustness on several real-world challenging imaging conditions compared to the state-of-the-art model and its predecessor designed for a leading manufacturer of endoscopy equipment. Currently, the developed model is evaluated in a multi-center clinical pilot study involving 200 patients.
Video-based characterization
A final contribution lies in the development of a lightweight CADx system for video-based characterization of BE neoplasia, improved for computational efficiency, temporal stability, and general clinical suitability for real-time decision support for doctors. This system is a complement to the developed CADe system used for initial lesion detection. It is developed as a compact, quantized algorithm supplemented by a temporal stability mechanism to improve prediction consistency on videos. The design enables accurate predictions with minimal computational cost, thereby facilitating potential integration as an embedded system within endoscopy systems. Validation in a clinical benchmarking study showed that the system not only outperformed general endoscopists and performs on-par with experts, but also enhances the diagnostic performance of general endoscopists, elevating their performance to expert level.
Enabling AI-assisted diagnostics
This research represents a major step toward making AI a routine part of GI endoscopy. By addressing the technical and operational challenges this research enables scalable, safe, and effective AI-assisted diagnostics. The work offers a blueprint for developing CADe/CADx systems that are not just smart, but also practical and clinically viable.
Title of PhD thesis: . Supervisors: Dr. Fons van der Sommen, Prof. Peter de With, and Prof. Jacques Bergman.