AutoCap: Automatic Anomaly Detection in Video Capsule Endoscopy

Almost, 3 million new gastrointestinal cancers occur every year and cause nearly 2 million deaths. Even for survivors, it largely influences the quality of life. Fortunately, it is largely possible to prevent these cancers by detecting and removing pre-cancerous lesions. A 1% increase in the detection of precursors in the colon during colonoscopy can decrease the risk of colorectal cancer with 3%. Screening on a regular basis is therefore recommended, and the Norwegian Government has decided to implement a national colon screening program for 55 years old citizens starting from 2019. Colonoscopy will be a cornerstone in this program, but the goal of population-wide screening will be hard to reach with the current personnel and equipment. Furthermore, some parts of the gastrointestinal system are hard to reach with traditional endoscopy examinations, like the small intestine. Moreover, existing procedures do not scale to the recommended screening program, the medical examinations introduce discomfort for patients and the cost per examination is high. In this respect, several manufacturers now provide video capsules (pills with small cameras) where the medical experts analyze the recorded video offline. This analysis, however, takes a long time and going fast forward increases the probability to overlook important anomalies. Therefore, automatic analysis and anomaly detection in capsule videos will be crucial for success. The goal of this project is to develop such a disease detection system, combining the latest video technologies with different artificial intelligence methods. The system should be able to detect multiple diseases with a high accuracy enabling a more efficient public health care system in order to reach the required scale and cost efficiency of future health care systems. We have developed tools for efficient annotation of video data, we have annotated a large number of images, and so far, our experimental results show what we can detect polyps with high accuracy in real-time. We have together with Vestre Viken collected a large annotated video datasett, and the next step is to refine our models based on this new data - both for traditional colonoscopy and capsule examinations. In the project, we have tested several models, but the datasets are made public so that researchers can both reproduce our results and develop new improved models.

First, the results have been moved to a new startup (Augere Medical AS) with the aim of making a product for GI video/image analysis. Thus, the results from the project will go directly into such a system. The system is scheduled to reach the European market in 2022 and the US market in 2023, and is expected to give a 100MNOK+ revenue from sales the first few years. Clinical trials performed on the system will show the real clinical effect, but if the system performs according to expectations, thousands of cancer cases will be prevented by the technology in the first few years of usage. Additionally, there have been developed an advanced tagging system to annotate medical data. This is used for all the collected data from OUS and Bærum Hospital. The knowledge derived from the research project has supplied invaluable insights into the significance of video quality for proper detection of the hardest-to-find lesions in the GI tract.

Gastrointestinal (GI) diseases are largely influencing quality of life with almost 3 million new GI cancers per year. The mortality is about 65%, but it is heavily influenced by early detection of neoplasia, where a 1% increase in detection can decrease the risk of cancer with 3%. For example, colon cancer is the third most common cause of cancer mortality, and it is a condition where early detection is important for survival, i.e., going from a low 10-30% 5-year survival probability if detected in later stages (III-IV) to a high 90% survival probability in early stages (I-II). Screening on a regular basis is therefore recommended, and the Norwegian Government has therefore decided to implement a national bowel screening program for 55 year old citizens starting from 2019 and providing full coverage by 2024. Colonoscopy will be a cornerstone in this program. This is a very important goal, but hard to reach with the current personnel and equipment. Some parts of the GI system is hard to reach with traditional endoscopy examinations like the small intestine, existing procedures does not properly scale to the recommended screening program, gastro- and colonoscopy examinations introduce discomfort for patients and the cost per examination is high. In this respect, several manufacturers now provide video capsules (pills) where the medical experts analyse the video of (parts of) the GI tract offline. This analysis, however, takes long time and going fast forward increase the probability to overlook important anomalies. Therefore, automatic analysis and anomaly detection in VCE videos will be crucial for success. The goal of this project is to develop such a system, combining latest video technologies with different types of artificial intelligence, able to detect multiple diseases with a high accuracy enabling a more efficient public health care system in order to reach the required scale and cost efficiency.