Artificial Intelligence for Classification of Soft-Tissue Masses at US.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Machine learning for diagnostic imaging

Background:

• Soft-tissue masses require accurate differentiation between benign and malignant types for appropriate patient management.
• Ultrasound (US) is a common imaging modality for evaluating soft-tissue masses, but interpretation can be challenging.
• Convolutional neural networks (CNNs) offer potential for automated image analysis and classification.

Purpose of the Study:

• To train CNN models for classifying benign versus malignant soft-tissue masses using US images.
• To develop a CNN model capable of differentiating between three common types of benign soft-tissue masses.
• To compare the diagnostic performance of the trained CNN models against experienced musculoskeletal radiologists.

Main Methods:

• Retrospective analysis of 419 patient US images with confirmed diagnoses.
• Training and validation of a CNN model using a modified VGG16 network on a Keras platform.
• Comparison of CNN model performance with two blinded musculoskeletal radiologists using McNemar test and Clopper-Pearson/DeLong methods for statistical analysis.

Main Results:

• The CNN model achieved 79% accuracy (AUC 0.91) in classifying malignant versus benign soft-tissue masses, comparable to expert radiologists.
• The model differentiating three common benign masses showed 71% accuracy.
• Performance metrics included accuracy, recall, specificity, and precision with estimated 95% confidence intervals.

Conclusions:

• Trained CNN models can effectively differentiate benign from malignant soft-tissue masses on US images.
• The CNN's performance in classifying malignant vs. benign masses matched that of experienced musculoskeletal radiologists.
• Further development may enhance the CNN's ability to differentiate specific types of benign masses.