A Gated Attention-Based Multiple Instance Learning and Test-Time Augmentation Approach for Diagnosing Active...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Disease Diagnosis
• Radiology and Diagnostic Imaging

Background:

• Axial spondyloarthritis (axSpA) is a chronic inflammatory disease affecting sacroiliac joints, necessitating early diagnosis to prevent irreversible damage.
• Conventional MRI interpretation for sacroiliitis is subjective, leading to inter-observer variability.
• Artificial intelligence (AI) offers potential solutions for objective and consistent diagnostic interpretation.

Purpose of the Study:

• To evaluate the diagnostic performance of a deep learning model based on Gated Attention Multiple Instance Learning (MIL) for automated sacroiliitis detection.
• To compare the proposed AI model's performance against a standard ResNet-18 architecture.
• To analyze the consistency of AI-driven findings with expert radiologist annotations.

Main Methods:

• A dataset of 554 subjects (276 axSpA patients, 278 controls) underwent MRI using axial T2-weighted fat-suppressed sequences.
• A deep learning model incorporating ResNet-18 feature extraction and a gated attention MIL framework was developed.
• Patient-wise data splitting for training, validation, and testing; Test-Time Augmentation (TTA) was used for enhanced robustness.

Main Results:

• The Gated Attention MIL model achieved 85.88% accuracy, 92.86% sensitivity, 79.07% specificity, and 86.67% F1-score on the independent test set.
• Attention heatmaps from the MIL module showed significant overlap with radiologist-annotated bone marrow edema.
• Test-Time Augmentation (TTA) improved overall classification accuracy by approximately 10%.

Conclusions:

• The Gated Attention MIL framework demonstrates high diagnostic performance for sacroiliitis detection.
• This AI approach serves as a valuable decision support tool for early axSpA diagnosis.
• Further validation on larger, multi-center datasets is recommended for clinical integration and generalizability.