Deep Learning-Based Prediction for Bone Cement Leakage During Percutaneous Kyphoplasty Using Preoperative Computed...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Spine Surgery

Background:

• Deep learning (DL) excels at medical image feature extraction, but models for predicting bone cement leakage (BCL) subtypes from preoperative images are lacking.
• Percutaneous kyphoplasty (PKP) is a common procedure for osteoporotic vertebral compression fractures, where BCL is a potential complication.
• Accurate prediction of BCL subtypes is crucial for optimizing surgical planning and patient outcomes.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for predicting bone cement leakage (BCL) subtypes during percutaneous kyphoplasty (PKP).
• To evaluate the model's effectiveness and generalizability using multicenter data.
• To compare the DL model's performance against spine surgeons in predicting BCL subtypes.

Main Methods:

• A retrospective study utilizing internal and external datasets for DL model development and testing.
• Integration of a vertebral segmentation module (3D U-Net) and a classification module (3D ResNet-50) for BCL subtype prediction.
• Performance evaluation using accuracy, area under the curve (AUC), sensitivity, and Cohen's kappa coefficient, comparing DL model with spine surgeons.

Main Results:

• The DL model achieved high vertebral segment identification accuracy (96.9%) and favorable AUC values (0.734-0.831) and sensitivities (0.649-0.900) on the internal dataset.
• The model demonstrated stable and generalizable performance on the external dataset with similar AUC values (0.709-0.818) and sensitivities (0.706-0.857).
• The DL model outperformed nonexpert spine surgeons in predicting BCL subtypes, with the exception of type II.

Conclusions:

• The developed DL model demonstrates satisfactory accuracy, reliability, generalizability, and interpretability in predicting BCL subtypes.
• The model's performance surpasses that of nonexpert spine surgeons, offering valuable insights for preoperative surgical decision-making in osteoporotic vertebral compression fractures.
• This AI-driven approach has the potential to enhance surgical planning and improve patient outcomes in PKP procedures.