Patch-Wise Approach with Vision Transformer for Detecting Implant Failure in Spinal Radiography

Area of Science:

• Orthopedics
• Radiology
• Artificial Intelligence

Background:

• Spinal implant failures are increasing due to more surgeries and an aging population.
• Early detection of spinal implant fractures on radiographs is challenging due to subtle signs and limited radiologist availability.
• Current diagnostic methods for spinal implant integrity often miss early signs of failure.

Purpose of the Study:

• To develop and validate a vision transformer-based deep learning model for detecting spinal implant fractures.
• To assess the impact of this AI tool on radiologist performance in identifying spinal implant failures.
• To evaluate the model's utility as a triage tool for spinal implant surveillance.

Main Methods:

• A DINOv2-based vision transformer model was trained on 9924 spinal radiographs to detect implants and analyze 224x224 pixel patches for fractures.
• The model was tested on 1538 images, and its performance was compared to three radiologists reviewing with and without AI assistance.
• Radiologist performance was measured using accuracy, F1 score, precision, recall, and generalized estimating equation (GEE) analysis.

Main Results:

• The AI model achieved a recall of 0.94, precision of 0.37, F1 score of 0.54, and accuracy of 0.83.
• Radiologist recall improved from 0.70 to 0.95 with AI assistance, with the most significant improvement in the least experienced reader (0.49 to 0.92).
• GEE analysis showed a significant diagnostic improvement (OR = 2.82, p < 0.001), especially for rod fractures.

Conclusions:

• The patch-wise transformer-based approach demonstrates high sensitivity for detecting spinal implant fractures.
• AI assistance significantly enhances radiologist diagnostic performance, particularly for less experienced readers.
• This AI model shows promise as an effective triage tool for spinal implant surveillance, improving early detection rates.