Deep learning in CT colonography: differentiating premalignant from benign colorectal polyps.

Area of Science:

• Radiology
• Artificial Intelligence
• Gastroenterology

Background:

• Colorectal cancer screening relies on detecting and characterizing polyps.
• Distinguishing premalignant from benign colorectal polyps is crucial for patient management.
• CT colonography is a non-invasive imaging technique for colorectal polyp detection.

Purpose of the Study:

• To investigate the efficacy of deep learning models in differentiating premalignant (adenoma) from benign (hyperplastic) colorectal polyps detected via CT colonography.
• To assess the ability of deep learning to visualize image regions critical for polyp classification.

Main Methods:

• A retrospective analysis of CT colonography images from an average-risk screening population.
• Two deep learning models, one with and one without polyp segmentation masks, were trained to classify polyps.
• Model performance was validated on an independent external multicenter test set using histopathology as the ground truth.

Main Results:

• The deep learning model with segmentation masks (SEG) achieved a ROC-AUC of 0.83, with 80% sensitivity and 69% specificity.
• The model without segmentation masks (noSEG) yielded a ROC-AUC of 0.75, with 80% sensitivity and 44% specificity.
• Deep learning models autonomously identified relevant polyp tissues for prediction, demonstrated by Grad-CAM++ visualization.

Conclusions:

• Deep learning offers a non-invasive method for differentiating premalignant from benign colorectal polyps on CT colonography.
• The developed deep learning approach has the potential to serve as an automated second reader, improving diagnostic accuracy.
• This technology could enhance patient selection for endoscopic polypectomy, particularly for polyps measuring 6-9 mm.