In our recent study, we developed an automated system for evaluating abdominal computed tomography (CT) scans to assist radiologists in managing their substantial workloads, thus improving patient outcomes. Our machine-learning model focuses on reliably identifying suspected bowel obstruction (BO) on abdominal CT scans.
We used an internal dataset of 1,345 annotated CT scans, of which only 670 were re-annotated by experienced radiologists, and an external dataset of 88 annotated CT scans. Our preprocessing pipeline included models to accurately locate and crop the abdomino-pelvic region. We then built, trained, and tested various neural network architectures for the binary classification (BO, yes/no).
The mixed convolutional network pretrained on a Kinetics 400 dataset achieved the best results. When sensitivity was prioritized, the model achieved a sensitivity of 1.00, a specificity of 0.84, and an F1 score of 0.88 with the internal dataset. The corresponding values for the external dataset were 0.98, 0.76, and 0.87.
Preprocessing is crucial and often overlooked. Constructing a pipeline capable of processing CT images holistically is a significant advancement. Triaging CT scans is not always useful in all hospital setups. We are working on more complex models for specific radiological questions in suspected bowel obstruction, such as recognizing functional obstruction and determining surgical needs.
A limitation in deploying our model, as with many medical imaging models, is the need for usability without requiring coding knowledge. Additionally, robust computational power and high-speed image transfer capabilities are essential.
Our 3D mixed convolutional neural network shows potential for automated BO classification, enhancing radiologist workflow.
Key points:
- Bowel obstruction’s rising incidence strains radiologists. AI can aid urgent CT readings.
- Employed 1345 CT scans, neural networks for bowel obstruction detection, achieving high accuracy and sensitivity on external testing.
- 3D mixed CNN automates CT reading prioritization effectively and speeds up bowel obstruction diagnosis.
Article: Deep learning for automatic bowel-obstruction identification on abdominal CT
Authors: Quentin Vanderbecq, Maxence Gelard, Jean-Christophe Pesquet, Mathilde Wagner, Lionel Arrive, Marc Zins & Emilie Chouzenoux
