With the use of computer-aided diagnostic systems, physicians can better diagnose and treat diseases in their patients. This has been shown to be especially useful for colorectal cancer (CRC), which is deadly and results in 900K deaths per year globally.
CRC originates in small, pre-cancerous lesions called polyps that are found deep inside the colon. These polyps can be identified and removed by a doctor with great success before they become cancerous and cause death.
Although there is no “perfect” way to detect colorectal polyps, colonoscopy has the advantage of being both a physical and visual examination. For this procedure to be successful in finding all benign or malignant tumors, though, it requires that medical personnel use their eyes as well as an instrument–the endoscope. This can sometimes lead to what’s known as incomplete diagnosis which means our doctor may not find anything if: 1) The tumor isn’t illuminated by the light source on the endoscopic tube; 2) If you have one small but deep-seated tumor so it resides just outside of view from where they are looking through at with lumen camera.
Google has published the latest research, “Detection of Elusive Polyps via a Large Scale AI System” in a potential solution to detecting polyps during colonoscopy using machine learning. Researchers have developed an AI model that can help medical staff accurately detect these abnormalities, and this system could potentially solve some of the issues with incomplete detection. The new research is based on Google’s previous work where they worked on maximizing scope for other areas of the body and preventing clinicians from missing parts that would otherwise go undetected by human eyes alone. In clinical trials following implementation, it was found that applying this technology had increased tumor detection rates significantly.
Google researchers have developed a system that is based on Convolutional Neural Networks. This new technology has two main advantages: it can solve problems with false-negative test results because polyps are difficult to find. Secondly, the false-positive rate of this system is very low, making it more likely for use in clinics.
Google researchers trained a system on 3600 procedures with 86 million frames and tested it on 1400 procedures (33 million frames). All the videos were de-identified. The 97% accurate detection of polyps, at 4.6 false alarms per procedure, is an improvement over previously published results. It is also exciting to note that this system does not care much about which neural network architecture it uses. The research team used two architectures: RetinaNet and LSTM-SSD, as models for testing and found them equally effective in their results!
As part of the research reported in this detection paper, researchers collaborated with a Jerusalem hospital to run 100 procedures using their system. The feedback from GIs was consistently positive, and it helped detect an average of one polyp per procedure that would have otherwise been missed by them while not missing any detected during these procedures.
Google Blog: https://ai.googleblog.com/2021/08/improved-detection-of-elusive-polyps.html