Intervertebral disc detection in X-ray images using faster R-CNN

Area of Science:

• Medical Imaging
• Deep Learning
• Spinal Biomechanics

Background:

• Ligament instability and injury affect 75% of motor vehicle accident patients.
• Accurate identification of osseous landmarks on spinal radiographs is essential for corrective calculations.
• Traditional methods for landmark identification are often inefficient and lack precision.

Purpose of the Study:

• To develop a deep learning-based object detection method for identifying landmark points in lateral lumbar X-ray images.
• To address the challenge of limited annotated medical datasets for deep learning applications.
• To improve the accuracy and efficiency of automatic osseous landmark identification in spinal imaging.

Main Methods:

• Fine-tuning a Faster-RCNN deep learning network, a state-of-the-art object detection model.
• Utilizing small annotated clinical datasets for training and fine-tuning.
• Comparing the deep learning approach with traditional sliding window detection methods.

Main Results:

• Achieved significantly better performance using only 81 lateral lumbar X-ray training images compared to traditional methods.
• Fine-tuned network with 974 images achieved an average precision of 0.905 on 108 test images.
• Demonstrated an average computation time of 3 seconds per image, outperforming traditional methods in accuracy and efficiency.

Conclusions:

• Deep learning, specifically fine-tuned Faster-RCNN, offers a highly accurate and efficient solution for automatic osseous landmark identification in spinal X-rays.
• The proposed method overcomes the challenge of limited annotated medical data by effectively utilizing small datasets.
• This approach has the potential to automate calculations for treating ligament instability and injury, improving patient outcomes.