SSGI04-04

Automated Segmentation and Worklist Prioritization of Pneumoperitoneum in Abdominal CT Images Using a Convolutional Neural Network

Wednesday, Dec. 2 8:30AM - 9:30AM Room: Channel 4



Participants
Robert J. Harris, PhD, Eden Prairie, MN (Presenter) Employee, Virtual Radiologic Corporation

For information about this presentation, contact:

robert.harris@vrad.com

PURPOSE

Pneumoperitoneum, the presence of free gas in the peritoneal cavity, can be a sign of critical pathology such as bowel perforation or trauma. Pneumoperitoneum is often diagnosed with abdominal CT and early detection is important to a patient's outcome. Our institution processes approximately 3,300 abdominal CT studies per day, of which 1.3% are positive for pneumoperitoneum. We hypothesized that a convolutional neural network could be trained to detect pneumoperitoneum in prospective patients in order to expedite patient care.

METHOD AND MATERIALS

Natural language processing (NLP) of radiology CT reports was used retrospectively to identify 297 body CT studies containing pneumoperitoneum. Axial CT images of these studies were annotated by a Board Certified radiologist to train a convolutional neural network. The training dataset consisted of 2,986 positive images and their segmentations, along with an equal number of negative images. A uNet model was trained using ResNet32 as the backbone. The model was first applied to a test cohort of 100 patients. This model was then integrated with our teleradiology pipeline to screen prospective patients for pneumoperitoneum in real-time, with NLP of the subsequent radiology report used as ground truth.

RESULTS

The model achieved an AUC of 0.906 on the test dataset. A detection threshold of 3 cc pneumoperitoneum was selected. Over a two-week period, for prospective patients, the model had a sensitivity of 50.1% and a specificity of 94.7%. The mean volume of pneumoperitoneum was 37.4 cc for true positives with a maximum of 413.5 cc.

CONCLUSION

An artificial intelligence model was trained to quantify pneumoperitoneum on CT images and implemented in a real-time clinical system. To our knowledge, this is the first use of machine learning to identify pneumoperitoneum on CT images and perform worklist prioritization for patients based on its presence. This model is currently being expanded to identify additional types of free air such as pneumothorax, pneumomediastinum, and soft tissue gas.

CLINICAL RELEVANCE/APPLICATION

Pneumoperitoneum is often an indicator of critical pathology and an artificial intelligence model can provide real-time detection of this condition on CT to save patients valuable time between scanning and treatment.

Printed on: 05/05/21