Assessment of automatic rib fracture detection on chest CT using a deep learning algorithm.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Rib fractures are common injuries in trauma patients.
• Accurate detection of rib fractures on CT scans is crucial for patient management.
• Current detection methods rely on human interpretation, which can be time-consuming and prone to error.

Purpose of the Study:

• To evaluate deep neural networks (DNNs) for automatic rib fracture detection on thoracic CT scans.
• To compare the performance of DNNs against attending-level radiologists.
• To assess the feasibility of using DNNs in a clinical setting for rib fracture diagnosis.

Main Methods:

• A retrospective study utilizing large internal (12,208 patients) and external (1613 patients) datasets of trauma patients with chest CT scans.
• Development of two cascaded deep neural networks based on an extended U-Net architecture for rib segmentation and fracture detection.
• Evaluation of model performance using metrics such as Area Under the Curve (AUC), sensitivity, and specificity, with comparisons to attending radiologist readings.

Main Results:

• The DNN model achieved high performance on the internal dataset with an AUC of 0.970, sensitivity of 93.3%, and specificity of 98.4%.
• On the external dataset, the model demonstrated strong performance with an AUC of 0.943, sensitivity of 86.2%, and specificity of 98.8%.
• The model's sensitivity (86.2%) significantly outperformed attending radiologists (70.5%) on the external dataset, while specificity was comparable (98.8%).

Conclusions:

• Deep learning models offer a feasible and effective approach for the automatic detection of rib fractures on chest CT scans.
• The performance of these deep learning algorithms is comparable to that of attending-level radiologists.
• Subtle rib fractures remain challenging, indicating a need for more extensive training data to improve detection of ambiguous lesions.