Fluoroscopic image-driven deep learning model for predicting intussusception irreducibility during air enema in...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pediatric Radiology

Background:

• Accurate identification of irreducible intussusception during air enema is critical for treatment.
• Current diagnostic methods rely on subjective interpretation, leading to inconsistencies.
• A deep learning (DL) framework was developed for objective prediction of irreducibility from fluoroscopic images.

Purpose of the Study:

• To develop and validate a DL framework for predicting irreducible intussusception.
• To objectively assess irreducibility using air enema fluoroscopic images.
• To compare the DL model's performance against human radiologists and existing techniques.

Main Methods:

• A hybrid ensemble DL model was trained on 770 irreducible and 1214 reducible intussusception cases.
• Model performance was evaluated on real-world and external test sets.
• Receiver operating characteristic (ROC) analysis and confusion matrix metrics were used for evaluation.

Main Results:

• The DL model achieved high diagnostic performance with AUCs of 0.89 and 0.883 on test sets.
• The model outperformed comparative methods and showed superior performance to an intermediate radiologist.
• As an assistive tool, the DL model significantly improved radiologists' diagnostic accuracy, increasing balanced accuracy and specificity.

Conclusions:

• The developed DL framework shows significant promise for identifying irreducible intussusception.
• This AI tool can serve as an effective decision-support system for radiologists during air enema procedures.
• The model has the potential to optimize enema strategies and improve patient outcomes.