Convolutional neural network (CNN) algorithms show promise for identifying gastrointestinal (GI) obstructions on x-ray in a pediatric population, researchers have reported.
The study findings could improve how clinicians determine treatment for kids presenting with GI symptoms, wrote a team led by Ercan Ayaz, MD, of the University of Health Sciences Türkiye in İstanbul. The group’s results were published February 28 in Diagnostic Interventional Radiology.
“Although some studies in the literature have shown the efficiency of CNN models in identifying small bowel obstruction with high accuracy for the adult population … our study is unique for the pediatric population and for evaluating the requirement of surgical versus medical treatment,” the investigators noted.
GI dilatations are commonly found in x-rays of pediatric patients who present in the emergency department with symptoms such as vomiting, pain, constipation, or diarrhea, the group explained. These patients need assessment as to whether there’s an obstruction that requires surgery; delayed diagnosis of this can lead to complications such as necrosis or perforation, which in turn can lead to death.
Ayaz and colleagues investigated the use of CNN models to distinguish between healthy children with normal intestinal gas identified on abdominal x-ray from those with GI dilatation or obstruction. They also sought to distinguish between patients with obstruction that would require surgery and those with other GI dilatations or intestinal blockages.
The researchers conducted a study that included 1,152 abdominal x-rays of patients with a surgical, clinical, and/or laboratory diagnosis of GI diseases with GI dilatation culled from their institution’s PACS archive. They created a control group that consisted of abdominal x-rays performed to detect abnormalities other than GI disorders. The images were categorized into three cohorts — surgically-corrected dilatation (n = 298), inflammatory/infectious dilatation (n = 314), and normal (n = 540) — and trained, validated, and tested five CNN models (ResNet50, InceptionResNetV2, Xception, EfficientNetV2L, and ConvNeXtXLarge).
The group found the following:
- For distinguishing between normal and abnormal images, ResNet50 had the highest accuracy rate (93.3%), with InceptionResNetV2 following second (90.6%).
- After automated cropping preprocessing, ConvNeXtXLarge demonstrated the highest accuracy rate for distinguishing between normal and abnormal images (96.9%), with ResNet50 (95.5%) and InceptionResNetV2 (95.5%) following second and third.
- EfficientNetV2L showed the highest accuracy for differentiating between surgically-corrected dilatation and inflammatory/infectious dilatation (94.6%).
The takeaway? CNN models such as these can be very helpful for assessing children presenting in the emergency department with GI symptoms, according to the authors.
“Deep-learning models can be integrated into [x-rays taken] in the emergency department as a decision support system [as they have] high accuracy rates in pediatric GI obstructions [and can immediately alert] the physicians about abnormal x-rays and possible etiologies,” they concluded.
The complete study can be found here.