Radiograph-based rheumatoid arthritis diagnosis via convolutional neural network

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Rheumatology

Background:

• Rheumatoid arthritis (RA) is a prevalent autoimmune disease with diagnostic challenges.
• Current diagnostic methods for RA can be subjective and error-prone.
• There is a critical need for automated and accurate RA detection systems.

Purpose of the Study:

• To develop a deep learning-based automated system for RA recognition and staging using radiographs.
• To assist physicians in rapid and precise diagnosis of rheumatoid arthritis.
• To explore the efficacy of Convolutional Neural Network (CNN) architectures for RA diagnosis.

Main Methods:

• A CNN-based automated diagnostic model was developed and evaluated.
• Five popular CNN architectures were explored for RA diagnosis from hand radiographs.
• The model was trained and tested on datasets of normal and RA-affected hand radiographs across five stages.

Main Results:

• CNN models demonstrated high performance in RA recognition, with AUCs above 90% and sensitivity over 98%.
• The GoogLeNet-based model achieved 97.80% AUC and 100.0% sensitivity for RA recognition.
• For RA staging, models achieved over 77% AUC and 80% sensitivity, with VGG16 showing 83.36% AUC and 92.67% sensitivity.

Conclusions:

• GoogLeNet and VGG16 models show superior performance for RA recognition and staging, respectively.
• The study confirms the feasibility and clinical applicability of CNNs for radiograph-based RA diagnosis.
• This automated system offers significant clinical value, particularly in resource-limited settings and for less experienced clinicians.