Development and validation of deep learning models for bowel obstruction on plain abdominal radiograph

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Bowel obstruction diagnosis relies on radiological images.
• Artificial intelligence (AI) offers potential for image analysis assistance.
• Deep learning (DL) models can be trained for specific diagnostic tasks.

Purpose of the Study:

• To develop and validate deep learning models for diagnosing bowel obstruction on abdominal radiographs.
• To compare the performance of different DL architectures (Xception, VGG16, ResNet).
• To evaluate the impact of an AI-aided diagnostic system on radiologist performance.

Main Methods:

• Retrospective collection of 2082 upright abdominal radiographs from four hospitals.
• Image labeling by three senior radiologists based on clinical outcomes within 48 hours.
• Development and validation of DL models, including Xception, VGG16, and ResNet, with gradient-weighted class activation mapping for explainability.

Main Results:

• The Xception model achieved the highest accuracy (0.863) in the validation set.
• In the test set, the Xception model (0.807 accuracy) outperformed other DL models and a junior radiologist (0.780).
• AI-aided diagnosis significantly improved radiologist accuracy to 0.887 (junior) and 0.913 (senior).

Conclusions:

• Validated DL computer vision models for diagnosing bowel obstruction on abdominal radiographs.
• DL-based computer-aided diagnostic systems can reduce practitioner workload.
• AI integration can enhance diagnostic accuracy in radiology.