Automated Risser Grade Assessment of Pelvic Bones Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Orthopedics

Background:

• Scoliosis management relies on accurate Risser grading from pelvic radiographs.
• Manual Risser grading is time-consuming and prone to inter-observer variability.
• Automating this assessment can enhance clinical workflow and diagnostic accuracy.

Purpose of the Study:

• To develop and validate deep learning models for automated Risser grade assessment.
• To utilize convolutional neural networks (CNNs) for analyzing pelvic radiographs.
• To improve the efficiency and reliability of Risser grading in adolescent scoliosis patients.

Main Methods:

• Trained two CNN models (left and right pelvis) on 1619 pelvic radiographs from patients aged 12-18 years.
• Employed a multimodal approach integrating image data with patient age and gender.
• Utilized Adam optimization, ReLU activation, dropout, batch normalization, and data augmentation to address class imbalance.

Main Results:

• The right pelvis CNN model achieved 83.64% accuracy, while the left model reached 80.56%.
• Both models demonstrated strong performance for Risser Grades 0, 2, and 4.
• The right model yielded a microaverage F1-score of 0.836 and ROC AUC of 0.895; the left model achieved 0.806 and 0.872, respectively.

Conclusions:

• CNN-based models offer an effective solution for automating Risser grade assessment.
• Automated grading can significantly reduce clinician workload and minimize assessment variability.
• This technology holds promise for streamlining scoliosis diagnosis and treatment planning.