XAOM

XAOM aligns and orients radiographs to standardize preprocessing of X-ray images for machine-learning-based diagnostic imaging.

Key Features:

• Two-stage pipeline: Implements distinct image alignment and image orientation stages for radiographic preprocessing.
• Alignment algorithms: Supports Hough transform, Fast line detection algorithm, and Principal Component Analysis (PCA) for image alignment.
• PCA performance: The PCA-based alignment achieved an average angular deviation of 1.65° and a median angular deviation of 0.11° on a test set of 200 pediatric X-ray images.
• Orientation classifiers: Uses adaptations of convolutional neural network topologies LeNet5, AlexNet, VGG16, VGG19, and ResNet50 for assigning orientation labels (north, south, east, west).
• VGG16 performance: The VGG16-based orientation network reached accuracy 0.993 on a test set of 4,221 images.
• Body-region coverage: Tunable to align and orient X-ray images from 21 different body regions.
• Standardization for ML: Provides standardized alignment and orientation of radiographs to support downstream machine learning analyses.

Scientific Applications:

• Fracture detection: Supplies consistently aligned and oriented radiographs as inputs to fracture detection algorithms.
• Disease region segmentation: Provides standardized inputs for segmentation methods that delineate disease-affected regions in radiographs.

Methodology:

XAOM performs two computational stages: image alignment using Hough transform, Fast line detection algorithm, and PCA (PCA reported avg angular deviation 1.65° and median 0.11° on 200 pediatric X-rays); and image orientation via adapted CNN topologies LeNet5, AlexNet, VGG16, VGG19, and ResNet50 (VGG16 reached 0.993 accuracy on 4,221 test images).