Application of an Automated Deep Learning Program to A Diagnostic Classification Model: Differentiating High-Risk...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Colorectal cancer screening relies on accurate polyp risk stratification.
• Distinguishing low-risk from high-risk adenomas is crucial for patient management.
• Standard white-light endoscopy requires skilled interpretation to assess polyp risk.

Purpose of the Study:

• To develop and evaluate a computer-aided diagnosis (CADx) model using automated deep learning (DL).
• To classify colorectal polyps ≤10mm into low- and high-risk categories.
• To compare the DL model's diagnostic performance against expert endoscopists and trainees.

Main Methods:

• A deep learning (DL) software (Neuro-T v3.2.1) was utilized for automated analysis.
• Still images of colorectal adenomas (≤10mm) were used for model training and validation.
• High-risk adenomas were defined by high-grade dysplasia or villous histology.

Main Results:

• The DL model achieved 93.8% accuracy, 81.0% precision, 85.7% recall, and 83.3% F1 score on an external validation dataset.
• The model's area under the receiver operating characteristic curve was 0.910 for high-risk and 0.914 for low-risk adenomas.
• Expert endoscopists achieved 95.1% accuracy, while trainees achieved 79.7% accuracy.

Conclusions:

• The automated DL-based CADx model demonstrates high diagnostic performance for risk stratification of colorectal polyps ≤10mm.
• The model's performance is comparable to expert endoscopists.
• The DL model significantly outperforms trainees in differentiating adenoma risk.