Classification and Visualisation of Normal and Abnormal Radiographs; A Comparison between Eleven Convolutional Neural...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate classification of radiographic images is crucial for diagnosing musculoskeletal conditions.
• Convolutional Neural Networks (CNNs) have shown promise in medical image analysis.
• Evaluating various CNN architectures is essential for optimizing diagnostic performance.

Purpose of the Study:

• To compare the performance of eleven distinct CNN architectures for classifying wrist radiographs.
• To assess the impact of data augmentation on the diagnostic accuracy of top-performing CNN models.
• To interpret the decision-making process of CNNs using Class Activation Mapping.

Main Methods:

• Eleven CNN architectures (GoogleNet, VGG-19, AlexNet, SqueezeNet, ResNet-18, Inception-v3, ResNet-50, VGG-16, ResNet-101, DenseNet-201, Inception-ResNet-v2) were employed.
• The models were trained and evaluated on the Stanford Musculoskeletal Radiographs (MURA) dataset for normal/abnormal classification.
• Performance was measured using accuracy and Cohen's kappa coefficient, with data augmentation applied to the best models.

Main Results:

• Inception-ResNet-v2 achieved the highest performance without augmentation (Mean accuracy = 0.723, Mean kappa = 0.506).
• Data augmentation significantly improved Inception-ResNet-v2 performance (Mean accuracy = 0.857, Mean kappa = 0.703).
• Class Activation Mapping provided insights into the network's focus on anomaly locations.

Conclusions:

• Inception-ResNet-v2 demonstrates strong potential for automated wrist radiograph classification.
• Data augmentation is a critical technique for enhancing CNN performance in this domain.
• CNN interpretability methods can aid in understanding diagnostic reasoning for medical imaging.