Automatic measurement of exophthalmos based orbital CT images using deep learning.

Objective, accurate, and efficient measurement of exophthalmos is imperative for diagnosing orbital diseases that cause abnormal degrees of exophthalmos (such as thyroid-related eye diseases) and for quantifying treatment effects. To address the limitations of existing clinical methods for measuring exophthalmos, such as poor reproducibility, low reliability, and subjectivity, we propose a method that uses deep learning and image processing techniques to measure the exophthalmos. The proposed method calculates two vertical distances; the distance from the apex of the anterior surface of the cornea to the highest protrusion point of the outer edge of the orbit in axial CT images and the distance from the apex of the anterior surface of the cornea to the highest protrusion point of the upper and lower outer edges of the orbit in sagittal CT images. Based on the dataset used, the results of the present method are in good agreement with those measured manually by clinicians, achieving a concordance correlation coefficient (CCC) of 0.9895 and an intraclass correlation coefficient (ICC) of 0.9698 on axial CT images while achieving a CCC of 0.9902 and an ICC of 0.9773 on sagittal CT images. In summary, our method can provide a fully automated measurement of the exophthalmos based on orbital CT images. The proposed method is reproducible, shows high accuracy and objectivity, aids in the diagnosis of relevant orbital diseases, and can quantify treatment effects.