Ankle Fracture Detection Utilizing a Convolutional Neural Network Ensemble Implemented with a Small Sample, De Novo...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology

Background:

• Deep learning models, particularly convolutional neural networks (CNNs), show promise in medical image analysis.
• Training CNNs de novo requires substantial datasets, which are often unavailable in specific medical domains.

Purpose of the Study:

• To investigate the feasibility of training CNN models de novo using a limited dataset for ankle radiographic analysis.
• To evaluate the impact of using single versus multiple radiographic views on model performance.
• To assess the effectiveness of model ensembling and data augmentation techniques.

Main Methods:

• Collected and processed 596 normal and abnormal ankle radiographic cases.
• Developed and trained Inception V3, Resnet, and Xception CNN models using Python and Tensorflow.
• Employed data augmentation during training.
• Evaluated model performance using single and three radiographic views.
• Created ensembles of trained CNN models and used a voting method for consolidated output.

Main Results:

• The ensemble of 5 CNN models achieved 76% accuracy using single radiographic views.
• Utilizing three radiographic views per case with the ensemble of all models resulted in the highest accuracy of 81%.
• The achieved 81% accuracy with a small dataset and de novo training is comparable to models trained on larger datasets or using pre-trained models.

Conclusions:

• De novo training of CNNs is viable for medical image analysis even with small datasets.
• Employing model ensembling and incorporating multiple radiographic views significantly enhances diagnostic accuracy.
• This approach offers a competitive alternative to methods requiring extensive datasets or manual feature engineering.