Development and Validation of a CT-based Diagnostic Models Using Artificial Intelligence for Detection of Small Bowel Obstruction

Overview

Small bowel obstruction (SBO) is a common non-traumatic surgical emergency. All guidelines recommend computed tomography (CT) as the first-line imaging test for patients with suspected SBO. The objectives of CT are multiple: (i) to confirm or refute the diagnosis of GI obstruction, defined as distension of the digestive tracts greater than 25 mm, and, when SBO is present, (ii) to confirm the mechanism (mechanical vs. functional), (iii) to localize the site of obstruction, i.e., the transition zone (TZ), (iv) to identify the cause, and (v) to look for complications such as strangulation or perforation, influencing management. Given the exponential increase in the number of scans being performed, especially in the setting of emergency management, methods to assist the radiologist would be useful to: 1. Sort the scans performed, allowing prioritization of the analysis of scans with a higher probability of pathology (occlusion in our case) 2. Help the radiologist to diagnose occlusion and its type (functional or mechanical), and to identify signs of severity. 3. To help the emergency physician and the digestive surgeon to make a decision on the management of the disease (surgical or medical). Machine learning has developed rapidly over the last decades, first thanks to the increase in data storage capacities, then thanks to the arrival of parallel processing hardware based on graphic processing units, in the context of radiological diagnostic assistance. Consequently, the number of studies on deep neural networks in medical imaging is increasing rapidly. However, few teams focus on SBO. The only published classification models have been produced for standard abdominal radiographs. No studies have used CT or 3D models, apart from our preliminary study on ZTs, despite the recognized advantages of CT for the diagnosis of SBO and the likely contribution of 3D models, which may be comparable to that of multiplanar reconstruction for the analysis of images in multiple planes of space.

Conditions