Objective:
It was Flowers in 1969 who first named the concaving grooves located at the border of the eyelid and cheek medially as “tear trough”. Multifactorial reasons causing orbital fat herniation, loss of the skin elasticity or the underlying support of this area will end up as a more significant medial periorbital hallowing, resulting in the tear trough deformity (TTD). A Versatile of the management and classification systems on TTD were then established. However, it was the diversity of the classification system that may bring into complexity during clinical use, especially for less experienced surgeons. The wide application of artificial intelligence (AI) technology in medicine and healthcare was believed to improve the present weakness of medical practice in particular extent such as time-consuming process requiring simple repetitive works, and to reduce inadvertent errors. The development on deep learning program had brough benefit in aiding pathological and imaging-based disease diagnosis. We then aim to establish a reliable and precise digital image grading system by AI deep learning program.

Material and Methods:
A total of 383 photos including periorbital area were enrolled in this study. The photos were categorized to according to the Barton’s classification by one senior plastic surgeon (H.C. Chen), one junior plastic surgeon (Y.T. Chen), and one junior surgical resident (S.S. Wu). The commercialized Medical AI Assistant (MAIA) software (Muen Biomedical and Optoelectronic Technologies Inc.) was then used for AI model establishment, with 344 photos (90%) for training, and 39 patients for testing. Confusion matrix table were demonstrated, and the result were then interpretated.

Results:
The results from the training group showed a good correlation of Cohen’s Kappa value 0.828(sensitivity of 90.86% and specificity of 66.8%). The heatmap for the zone of interest of AI training group demonstrated a clear focus on the periorbital area, mostly on the lower eyelid, which was compatible to our area of interests. As for the testing group, the kappa value was 56.5%(sensitivity 40.83%, specificity 84.47%). The heatmap in the testing group also showed a precise focusing on the periorbital area.

Conclusion:
The grading for the tear trough deformity can be accurately categorized under the assistance of AI deep learning technology via easily accessible images, as a trustable tool for aiding the clinician on determining the subsequent treatment for the patients.

Download