Artificial intelligence techniques for the automatic detection of colorectal polyps

Abstract

Colorectal cancer (CRC) is one of the most prevalent malignant tumors worldwide. These neoplasms originate from adenomatous lesions or polyps that must be resected to prevent the development of the disease, and that can be done through a colonoscopy. Polyps are reported during colonoscopy in 40% of men and 30% of women (hyperplastic, adenomatous, serrated, among others), and, on average 25% are adenomatous polyps (the main indicator of quality in colonoscopy). However, these lesions are not easy to visualize because of the multiplicity of blind spots in the colon and human errors associated with the performance of the procedure. Several research works have reported that about 25% of colorectal polyps are overlooked or undetected during colonoscopy, and as a result, the patient may have interval cancer. These figures show the need for a second observer (artificial intelligence system) to reduce the possibility of not detecting polyps and prevent colon cancer as much as possible.

Objective: To create a computational method for the automatic detection of colorectal polyps using artificial intelligence using recorded videos of colonoscopy procedures.

Methodology: Public databases of colorectal polyps and a data collection constructed in a university hospital were used. Initially, all the frames in the videos were normalized to reduce the high variability between databases. Subsequently, polyps were detected using a deep learning method with a convolutional neural network. This network starts with weights learned from millions of natural images taken from the ImageNET database. Network weights are updated using colonoscopy images, following the fine-tuning technique. Finally, polyps are detected by assigning each box a probability of polyp presence and determining the threshold that defines when the polyp is present in a box.

Results: This approach was trained and evaluated with 1 875 cases collected from 5 public databases and the one built in the university hospital, which total approximately 123 046 frames. The results obtained were compared with the markings of different experts in colonoscopy, obtaining 0.77 accuracy, 0.89 sensitivity, 0.71 specificity, and a receiver operating characteristic curve of 0.87.

Conclusion: This method detected polyps in an outstanding way, overcoming the high variability caused by the types of lesions and bowel lumen condition (loops, folds or retractions) and obtaining a very high sensitivity compared with an experienced gastroenterologist. This may help reduce the incidence of human error, as it is one of the main factors that cause polyps to not be detected or overlooked during a colonoscopy.