Deep convolutional neural network for rib fracture recognition on chest radiographs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Rib fractures are common in trauma, with missed diagnoses increasing patient morbidity and mortality.
• Current imaging methods like CT scans are costly and involve radiation; ultrasound is operator-dependent.
• There is a need for accurate, accessible, and efficient methods for rib fracture detection.

Purpose of the Study:

• To evaluate the effectiveness of deep convolutional neural networks (DCNNs) for identifying rib fractures on standard chest radiographs.
• To explore DCNNs as a practical alternative to existing diagnostic tools for rib fractures.

Main Methods:

• A retrospective dataset of chest radiographs with documented rib fractures was compiled.
• Deep learning models (DLMs) were trained on 2000 region-of-interest (ROI) slices per category (fractured, non-fractured, background).
• Images were segmented to 128x128 pixels for optimized DCNN training.

Main Results:

• Several DCNN models achieved high validation accuracies: AlexNet (92.6%), GoogLeNet (92.2%), EfficientNetb3 (92.3%), DenseNet201 (92.4%), and MobileNetV2 (91.2%).
• These results indicate strong performance of DCNNs in distinguishing fractured from non-fractured ribs.

Conclusions:

• Integrating DCNNs into clinical decision support systems can significantly reduce missed rib fracture diagnoses.
• This AI-driven approach promises to decrease morbidity and mortality in trauma patients.
• DCNNs offer a practical and effective solution for improving chest trauma diagnosis and patient care.