Colorectal cancer is a major health problem. Early diagnosis can be established by a systematic endoscopic screening of the whole target population. The scale of the systematic screening increases the burden on pathology laboratories, especially in combination with the limited available time frame to process the samples.
Whole slide scanners can provide whole slide scans (WSI) of classical glass slides. The WSI can be analyzed manually by a pathologist or automatically by using "Digital image recognition techniques". These techniques have been improved greatly thanks to advances in deep learning and "Artificial Neural Networks" can be trained to recognize patterns in digital images with high accuracy.
The aim of the study is to design a model, based on machine learning techniques, that recognizes major zones of interest: normal tissue, low grade- and high grade dysplasia and invasive carcinoma on WSI of colon biopsies.
In a pilot study we were able to build a model identifying abnormalities with an accuracy of 95% based on 190 digital images. In a follow up study we examined 73 WSI of colonic polyps. Annotation of zones of interest were made independently by 4 gastro-intestinal pathologists. Based on these data we designed and trained a machine learning (ML) model that recognizes and classifies accurately zones of interest focusing on pattern recognition and not on object (cell) detection/classification.
This model can be used as a pre diagnostic tool that offers support to the pathologist and will help to provide an accurate pathology report within a short time.