Deep Learning-Enabled Detection of Pneumoperitoneum in Supine and Erect Abdominal Radiography: Modeling Using...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Detecting pneumoperitoneum on abdominal X-rays, especially when patients are supine, presents diagnostic challenges.
• Deep learning offers a potential solution to improve the accuracy and efficiency of pneumoperitoneum detection.

Purpose of the Study:

• To develop and externally validate a deep learning model for detecting pneumoperitoneum using both supine and erect abdominal radiography.
• To assess the model's performance compared to human radiologists.

Main Methods:

• A deep learning model was developed using knowledge distillation and a semi-supervised learning method (DISTL) leveraging Vision Transformer.
• The model was pre-trained on chest radiographs, then fine-tuned and self-trained on labeled and unlabeled abdominal radiographs.
• Internal and external validation was performed on multiple datasets, evaluating performance using the area under the receiver operating characteristic curve (AUC).

Main Results:

• The model achieved high AUCs in internal validation (up to 0.968 for erect, 0.881 for supine).
• External validation showed strong performance with AUCs up to 0.944 (erect) and 0.852 (supine).
• Radiologists' diagnostic performance improved when assisted by the deep learning model.

Conclusions:

• The developed deep learning model, trained with the DISTL method, accurately detects pneumoperitoneum on abdominal radiography.
• The model demonstrates effectiveness in both supine and erect patient positions, offering a valuable tool for clinical practice.