Deep learning based CT grading system for sacroiliitis: a multi-center studydemonstrating superior accuracy and...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Rheumatology
• Deep Learning for Disease Diagnosis

Background:

• Sacroiliitis is a key indicator of axial spondyloarthritis (axSpA).
• Accurate grading of sacroiliitis on CT images is crucial for axSpA diagnosis and management.
• Current grading methods can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (DCNN) for automated sacroiliitis grading.
• To assess the diagnostic accuracy and efficiency of the DCNN compared to human readers.
• To evaluate the model's performance on internal and external validation datasets.

Main Methods:

• A 3D-ResNet50 DCNN model was developed using CT images from 1341 axSpA patients.
• The model was validated on internal (130 patients) and external (249 patients) datasets.
• Diagnostic sensitivity and reading time were compared between the model and six human readers (rheumatologists and radiologists).

Main Results:

• The 3D-ResNet50 model achieved high diagnostic accuracies (87.1%-89.9%) across training, testing, and validation sets.
• The model's area under the curve (AUC) values in external validation exceeded those of human readers, particularly for Grade II sacroiliitis.
• The DCNN significantly reduced diagnostic time per case (2.74s) compared to human readers (119.4s).

Conclusions:

• The developed 3D-ResNet50 model demonstrates high accuracy and efficiency in grading sacroiliitis.
• The model's performance surpasses that of rheumatologists and radiologists, suggesting its potential clinical utility.
• Integration of this AI tool could assist in the diagnosis and management of axSpA.