Effects of a comprehensive brain computed tomography deep learning model on radiologist detection accuracy

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Machine Learning Applications in Healthcare

Background:

• Non-contrast computed tomography of the brain (NCCTB) is crucial for detecting intracranial pathology.
• Interpretation of NCCTB scans can be prone to errors, impacting diagnostic accuracy.
• Machine learning (ML) offers potential to augment clinical decision-making in radiology.

Purpose of the Study:

• To assess the performance of a deep learning (DL) model in assisting radiologists with NCCTB interpretation.
• To compare the diagnostic accuracy of radiologists with and without DL model assistance.
• To evaluate the impact of DL assistance on radiologist interpretation time.

Main Methods:

• A DL model was trained on a large dataset (212,484 scans) of NCCTB.
• Thirty-two radiologists reviewed 2848 NCCTB scans, both with and without DL model assistance.
• Performance was evaluated using area under the receiver operating characteristic curve (AUC) and Matthews correlation coefficient (MCC) against a gold standard.

Main Results:

• The DL model achieved an average AUC of 0.93 across 144 findings.
• Radiologist performance significantly improved when assisted by the DL model (average AUC 0.79 vs. 0.73 unassisted).
• DL assistance led to significantly improved AUC for 91 findings and a significant reduction in reading time.

Conclusions:

• A comprehensive deep learning system significantly enhances radiologist accuracy in detecting a wide range of abnormalities on NCCTB scans.
• The DL model demonstrated strong standalone performance and improved radiologist interpretation efficiency.
• This technology holds potential to reduce diagnostic errors, improve workflow efficiency, and facilitate timely patient care.