Hybrid deep learning ensemble model for detecting small to medium rotator cuff tears from shoulder radiographs

Area of Science:

• Orthopedic imaging
• Artificial intelligence in medicine
• Radiology

Background:

• Rotator cuff tears (RCTs) are common orthopedic conditions requiring advanced imaging for diagnosis.
• Plain radiography may not reliably differentiate small to medium RCTs from other shoulder pathologies.
• Deep learning (DL) offers potential for preliminary RCT assessment, reducing unnecessary advanced imaging and expediting care.

Purpose of the Study:

• To develop and validate a hybrid deep learning ensemble model for detecting small to medium rotator cuff tears (RCTs) from shoulder radiographs.
• To assess the model's performance in classifying rotator cuff images compared to individual deep learning models.

Main Methods:

• A retrospective study utilized 587 shoulder radiographs, divided into training, validation, and independent test sets.
• Three pre-trained deep learning models (ResNet-50, ResNet-101, InceptionV3) were fine-tuned using transfer learning.
• A hybrid DL ensemble model was created using a majority voting strategy for improved classification accuracy.

Main Results:

• The hybrid DL ensemble model achieved high performance on the independent test set: 81% accuracy, 72% precision, 100% recall, 63% specificity, and an AUROC of 0.93.
• On the validation set, the model demonstrated 78% accuracy, 79% precision, 87% recall, 66% specificity, and an AUROC of 0.85.
• The hybrid model outperformed individual DL models in classifying rotator cuff images.

Conclusions:

• The developed hybrid DL ensemble model effectively detects small to medium RCTs from shoulder radiographs.
• This AI-driven approach serves as a cost-effective and efficient tool for preliminary RCT assessment.
• The model holds potential to enhance diagnostic reliability and support clinical decision-making as a screening tool.