Machine Learning Approach Helps to Improve the Diagnosis of Colon Cancer

Researchers are developing a new imaging process that will provide an accurate computer-aided diagnosis of colon cancer in real-time. The team used deep learning, which is a type of machine learning on frames of imaging data from tissue samples to find out the accuracy of the method.

During the pilot study;

• The team was able to find cancer with 100% accuracy through PR-OCT imaging. They combined it with machine learning to help differentiate healthy tissue from cancerous tissue and precancerous polyps.

• Flexible colonoscopy is the current screening standard for colorectal cancer. It involves visually inspecting the mucosal lining of the rectum and colon using a camera that is mounted on an endoscope. Any abnormality areas found are biopsied for analysis.

• But colonoscopy has drawbacks too – it depends on visual detection and may not identify small lesions and other changes that are in deeper layers of the bowel wall. This makes it challenging to treat early malignancies.

• PR-OCT imaging is based on optical coherence tomography, which is an imaging technology used in ophthalmology to capture images of the retina. It has been advanced for other purposes to offer high spatial and depth resolutions of up to 1 – 2 mm imaging depth.

• The technique can identify any differences in the way healthy and cancerous tissue refract light. It is also highly sensitive to early and precancerous morphological changes. In the future, it can be advanced and used for real-time noninvasive imaging together with colonoscopy endoscope to screen precancerous polyps that are situated deeper in the organ and other early-stage colorectal cancers.

• Colorectal tissue forms a pattern on OCT imaging caused by the light attenuation of healthy mucosa microstructures. It can detect a structural pattern within images and help to classify both abnormal and cancerous tissue to ensure an accurate diagnosis.

The research team is now designing a catheter to help analyze the pattern formed on the surface of colon tissue to provide a score of cancer probability. By developing the computation speed and the catheter, the technique can give feedback in four seconds.