A Convolutional Neural Network Deep Learning Model Trained on CD Ulcers Images Accurately Identifies NSAID Ulcers

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) for video capsule endoscopy (VCE) is an emerging field for detecting gastrointestinal conditions.
• Non-steroidal anti-inflammatory drugs (NSAIDs) can cause small bowel ulcers, a key differential diagnosis for Crohn's disease (CD) ulcers.
• Evaluating DL performance on NSAID ulcers is crucial for accurate VCE analysis.

Purpose of the Study:

• To assess the efficacy of a DL network trained on CD ulcers for identifying NSAID-induced ulcers in VCE.
• To compare the diagnostic performance of the DL network for NSAID ulcers versus CD ulcers.

Main Methods:

• A DL network was trained using 17,640 VCE images of CD ulcers and normal mucosa.
• The network's performance was evaluated on a separate dataset of 1,605 VCE images of NSAID-induced enteropathy.
• Area Under the Receiver Operating Curve (AUC) was the primary performance metric.

Main Results:

• The DL network achieved an AUC of 0.97 for identifying NSAID mucosal ulcers.
• The diagnostic accuracy for NSAID ulcers was comparable to that for CD ulcers (AUC 0.94-0.99).
• The DL network demonstrated high performance in detecting both types of small bowel ulcers.

Conclusions:

• A DL network trained on VCE images of CD ulcers can effectively identify NSAID-induced ulcers.
• This highlights the potential for generalizable DL applications in VCE for small bowel ulcer detection.
• Future VCE DL tool development should consider this cross-diagnostic capability.