Application of an Automated Deep Learning Program to A Diagnostic Classification Model: Differentiating High-Risk Adenomas Among Colorectal Polyps 10 mm or Smaller

Objective

This study aimed to develop a computer-aided diagnosis (CADx) model using an automated deep learning (DL) program to classify low- and high-risk adenomas among colorectal polyps ≤ 10 mm with standard white-light endoscopy.

Methods

Still images of colorectal adenomas ≤ 10 mm were extracted. High-risk adenomas were defined as high-grade dysplasia or adenomas with villous histology. Neuro-T version 3.2.1 (Neurocle Inc., Seoul, Republic of Korea), an automated DL software, was used for DL. Accuracy, precision, recall, and F1 score of the DL model were calculated. Endoscopy experts and trainees were invited to diagnose endoscopic images to compare their diagnostic accuracy with that of the DL model.

Results

A total of 2696 endoscopic images (2460 images of low-grade and 236 of high-grade adenomas) were used for training the DL model. In classifying high- and low-risk adenomas in the external validation dataset (398 images of low-grade and 41 images of high-grade adenomas), the model demonstrated 93.8% accuracy, 81.0% precision, 85.7% recall, and 83.3% F1 score overall. The area under the receiver operating characteristic curve for classifying high- and low-risk adenomas was 0.910 and 0.914, respectively. The expert endoscopists and trainees showed an overall accuracy of 95.1% and 79.7%, respectively, for discriminating high- and low-risk adenomas in the external validation dataset.

Conclusions

The CADx model established by the automated DL program showed high diagnostic performance in differentiating high- and low-risk adenomas among colorectal polyps ≤ 10 mm. The performance of the model was comparable to the experts and superior to the trainees.