APPLICATION OF DEEP LEARNING NEURAL NETWORK FOR HISTOLOGICAL PREDICTION OF COLON POLYP IMAGES WITH BLI ZOOM TECHNOLOGY

Aims:

In our present study we aimed to develop an Artificial Intelligence-based Decision Support System (AI-DSS) that can be used to analyze the polyp images in differentiation between neoplastic and non-neoplastic subcentrimetric polyps.

Methods:

We enrolled 755 HD images with Blue Light Imaging (BLI) zoom technology of total 334 histologically identified colorectal polyps. We set up 4 subgroups for training and testing with deep learning: A: training and testing set data was selected only from typical polyps, B: training set is made from only typical polyps, and test set is made from typical and atypical polyps, C: training set is made from only typical polyps, and test set is made randomly from the whole set (mixed typical and atypical), D: both train and test set are made of polyps randomly selected from the whole set. Images for the test sets were selected randomly following these criteria. Images from the same polyp were not selected to both train and test set.

Results:

The images went through a pre-process algorithm, and then we trained and tested the neural network. We also assessed which training parameters gave the best test results. The test groups had the following accuracy, sensitivity, specificity, PPV and NPV values to predict adenomatosus polyps as follows: Group A: 95%, 96.7%, 93.3%, 93.5%, 96.6%; Group B: 73.6%, 76.5%, 68.4%, 81.3%, 61.9%; Group C: 89.4%, 91.5%, 87.2%, 87.8%, 91.1%; Group D: 73,1%, 76,9%, 69.2%, 71.4%, 75%, respectively.

Conclusions:

This AI-DSS is able to predict the polyp histology with high accuracy, if the neural network is trained on typical images. Accuracy of the algorithm could be further increased with higher number of collected images. Application of Deep Learning Neural Network with BLI zoom virtual-chromoendoscopy provide a potential for real-time endoscopic optical diagnosis of hyperplastic polyps to support resect and discard strategy.