Deep Learning-Based 2.5D Multi-Window CT Image Recognition Platform for Thyroid-Associated Ophthalmopathy Diagnosis

Bowen Yang

Co-Presenters: Individual Presentation

College: Hennings College of Science Mathematics and Technology

Major: BS.COMPUTER/SCI

Faculty Research Mentor: Kuan Huang

Abstract:

Eye CT images play a central role in the clinical diagnosis of thyroid-associated ophthalmopathy (TAO) because they provide detailed visualization of orbital structures, including extraocular muscles, orbital fat, and globe position, which are essential for assessing proptosis and tissue abnormalities. However, interpreting CT images requires substantial medical knowledge of orbital anatomy, longitudinal comparison across time points, and subjective clinical judgment. This process is labor-intensive, operator-dependent, and may overlook subtle image features. This study aims to develop a deep learning–based eye CT image recognition platform to automatically extract key TAO-related findings from CT scans, such as the presence or absence of proptosis, and to assist radiologists in report generation and clinical decision support.
Our approach includes two crucial innovations: Firstly, we introduce an unsupervised algorithm to tackle the problem of analyzing whole-brain CT data. This algorithm works directly on the original DICOM CT data and automatically extracts key slices containing the eyeballs, essentially removing the interference of non-relevant whole-brain structures. Secondly, we create a multi-modal input process that is an image + text, which integrates the clinical manifestations of the patient with the image characteristics that imitate the overall diagnostic reasoning of radiologists.
This system enables automatic extraction of key features of thyroid-associated ophthalmopathy in CT images, including exophthalmos and extraocular muscle enlargement. The extracted findings are compiled into structured diagnostic reports using a large language model with predefined templates, providing objective and quantitative support for clinical decision-making.