Neural network application for assessing thyroid-associated orbitopathy activity using orbital computed tomography.

Area of Science:

• Ophthalmology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Thyroid-associated orbitopathy (TAO) is an autoimmune condition affecting the eye socket.
• Accurate assessment of TAO activity is crucial for timely and effective treatment.
• Current diagnostic methods may lack precision in differentiating active from inactive disease stages.

Purpose of the Study:

• To develop and validate a novel neural network (NN)-based method for assessing TAO patient activity.
• To utilize orbital computed tomography (CT) scans and clinical data for TAO activity evaluation.
• To compare the performance of the proposed NN with existing deep learning models.

Main Methods:

• Orbital CT scans from 144 active and 288 inactive TAO patients were analyzed.
• A preprocessing pipeline involved selecting eleven key image slices and segmenting regions of interest.
• A custom NN was designed, integrating data from 13 information extraction pipelines and patient age/sex data.

Main Results:

• The proposed NN achieved an area under the receiver operating curve (AUROC) of 0.871, with 0.786 sensitivity and 0.779 specificity.
• Performance significantly surpassed comparison models CSPDenseNet (AUROC 0.819) and ConvNeXt (AUROC 0.774).
• Ablation studies indicated optimal NN performance with three to five active information pipelines.

Conclusions:

• The developed NN-based method demonstrates high potential as a reliable tool for diagnosing TAO activity.
• The findings suggest that integrating multi-pipeline information extraction enhances diagnostic accuracy.
• Further validation is recommended to establish this method as a clinical standard for TAO assessment.