Our automated reading capsule endoscopy technology is based on image processing algorithms developed at the ETIS research laboratory (CY Cergy Paris University, ENSEA, CNRS) in collaboration with the Saint-Antoine University Hospital (Sorbonne University and Assistance Publique - Hôpitaux de Paris).
The architecture of our algorithms is designed to be integrated into a Software as a Service (SaaS) environment.
This technological development is built on several academic research projects conducted since 2014 with the support of SATT ERGANEO, ENSEA and the Initiative d’Excellence Paris Seine.
Our solution is at the interface between Machine Learning and intelligent embedded systems. The development of our algorithms is based on the most recent Deep Learning tools and, in particular, the use of original architectures in the form of "convolutional neural networks". The major challenge lies in developing our capacity to use these architectures with limited computing resources. Enabling their real-time use in the context of videocoloscopy, for example, and within SaaS-compatible timeframes in the context of capsule endoscopy.
Our algorithm has been developed using Deep Learning techniques. It is built on the latest approaches in which the architectures and convergence criteria ensure an optimal balance between performance and computing time.
The architectures evolve in accordance with the different tasks required: altogether, they offer a flexible use of the learning database and answer to the expectations of physicians.
AUTOMATED DETECTION OF IMAGES OF INTEREST
Vascular or inflammatory abnormalities, polyps, presence of blood (fresh clots or melena)
CHARACTERIZATION AND RELEVANCE OF LESIONS DETECTED
Angiectasia, ulcers, polyps
QUALITY OF THE ENDOSCOPIC EXAMINATION
Completeness of the examination (complete visualization of the digestive tract segment under analysis)
Adequate or inadequate quality of bowel preparation (visualization of the intestinal mucosa)
The choice of the algorithmic architectures is also informed by the dynamic nature of the database: the expansion of the database facilitates continuous learning. The database is designed to evolve and to be updated in accordance with the clinical cases that have been analysed. This "active learning" approach leads to continuous improvement in performance and, if necessary, to the consideration of new classes of lesions (in accordance with their nature or their position before or after the small bowel, for example).
A DATABASE OF OVER 4,000 RECORDINGS MADE BY THE LATEST GENERATION OF CAPSULE ENDOSCOPY DEVICES
INITIAL EVALUATIONS CARRIED OUT ON 140 VIDEOS (1,750,000 IMAGES) SHOW EXCELLENT DIAGNOSTIC PERFORMANCE IN LESS THAN 5 MINUTES