Predicting Perceived Reporting Complexity of Abdominopelvic Computed Tomography With Deep Learning

Area of Science:

• Radiology
• Medical Imaging Analysis
• Artificial Intelligence in Medicine

Background:

• Assessing reporting complexity in radiology is crucial for workflow optimization.
• Computed tomography (CT) studies of the abdomen and pelvis are common and vary in complexity.

Purpose of the Study:

• To compare human and automated estimates of reporting complexity for abdominal and pelvic CT scans.
• To evaluate the feasibility of using AI for complexity assessment in radiology.

Main Methods:

• 1019 CT studies were categorized into 3 complexity levels by 3 radiologists.
• A two-stage neural network model was trained to predict study complexity.
• Radiologist consensus served as the ground truth.

Main Results:

• High agreement (99%) was observed among radiologists when at least two out of three agreed.
• The neural network model achieved 55% agreement with the radiologist consensus.
• Most model errors (90%) involved a one-level difference in complexity prediction.

Conclusions:

• Radiologist agreement on CT reporting complexity is moderate.
• Automated prediction of reporting complexity shows promise as a tool for radiology practice analytics.