ACCURACY OF ARTIFICAL INTELLIGENCE BASED DECISION SUPPORT SYSTEM COMBINED WITH BASIC CLASSIFICATION IN COLON POLYP DETERMINATION

Aims:

All colon polyps have to be removed for histological analysis due to regulatory reasons, causing real health care burden. Virtual-chromoendoscopy only in expert hands may be used for real-time diagnosis of polyp characteristics to support resect and discard strategy. In our present study, we aimed to develop an Artificial Intelligence-based Decision Support System (AI-DSS) that can assist the decision of endoscopist in the real-time determination of sub-centimetric polyp's histology with high accuracy.

Methods:

We enrolled 334 histologically identified colon polyp, having at least one good quality, zoomed and non-zoomed HD image with Blue Light Imaging (BLI) virtual-chromoendoscopy. The images were characterized by an expert endoscopist using BASIC classification, including description of surface characteristics, (pseudo)depression, pit pattern and vessel structure of the polyp (2 – 2-2 – 4-3 – 4 variable options). We randomly generated subgroups from polyps to train the multilayered deep learning neural network (test set: 100 hyperplastic/neoplastic (50 – 50%), train set: 234 polyps), then run the training process four-times with different subgroups and two output classes: hyperplastic and neoplastic. The decision is based on the highest softmax value of the output neurons. The training ran for several hundreds of epochs, and we stopped it when the test accuracy reached the best results to prevent overfitting on the training set. We analyzed the result with/without including size and localisation of the polyp (2 – 3 options).

Results:

The maximum accuracy of different training tests were 91, 94.5, 93.75, 88%/87.75, 92.5, 91.5, 88.5% separately with a total 91.81% and 90.06%, with/without including localisation and size information.

Conclusions:

AI-DSS combined with BASIC classification is able to predict the polyp histological dignity with a high accuracy, which could be further increased with higher number of images. This software can support the everyday clinical decision of resect and discard strategy during polypectomy, moreover it can be used by trainee endoscopists.