Ju-Yi Hung1, Chandrashan Perera2, Ke-Wei Chen3, David Myung2, Hsu-Kuang Chiu4, Chiou-Shann Fuh5, Cherng-Ru Hsu6, Shu-Lang Liao7, Andrea Lora Kossler8. 1. Ophthalmology, Byers Eye Institute, Stanford University School of Medicine, Palo Alto, California, United States; Ophthalmology, Taipei Medical University Hospital, Taipei, Taiwan; Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan. 2. Ophthalmology, Byers Eye Institute, Stanford University School of Medicine, Palo Alto, California, United States. 3. Ophthalmology, Byers Eye Institute, Stanford University School of Medicine, Palo Alto, California, United States; Biomedical Engineering, National Cheng Kung University, Tainan, Taiwan. 4. Computer Science, Stanford University, Stanford, California, United States. 5. Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan. 6. Ophthalmology, National Taiwan University Hospital, Taipei, Taiwan; Ophthalmology, Tri-Service General Hospital, National Defense Medical Center, Taipei, Taiwan. 7. Ophthalmology, National Taiwan University Hospital, Taipei, Taiwan; College of Medicine, National Taiwan University, Taipei, Taiwan. Electronic address: liaosl89@ntu.edu.tw. 8. Ophthalmology, Byers Eye Institute, Stanford University School of Medicine, Palo Alto, California, United States. Electronic address: akossler@stanford.edu.
Abstract
PURPOSE: Blepharoptosis is a known cause of reversible vision loss. Accurate assessment can be difficult, especially amongst non-specialists. Existing automated techniques disrupt clinical workflow by requiring user input, or placement of reference markers. Neural networks are known to be effective in image classification tasks. We aim to develop an algorithm that can accurately identify blepharoptosis from a clinical photo. METHODS: A total of 500 clinical photographs from patients with and without blepharoptosis were sourced from a tertiary ophthalmic center in Taiwan. Images were labeled by two oculoplastic surgeons, with an independent third oculoplastic surgeon to adjudicate disagreements. These images were used to train a series of convolutional neural networks (CNNs) to ascertain the best CNN architecture for this particular task. RESULTS: Of the models that trained on the dataset, most were able to identify ptosis images with reasonable accuracy. We found the best performing model to use the DenseNet121 architecture without pre-training which achieved a sensitivity of 90.1 % with a specificity of 82.4 %, compared to the worst performing model which was used a Resnet34 architecture with pre-training, achieving a sensitivity of 74.1 %, and specificity of 63.6 %. Models with and without pre-training performed similarly (mean accuracy 82.6 % vs. 85.8 % respectively, p = 0.06), though models with pre-training took less time to train (1-minute vs. 16 min, p < 0.01). CONCLUSIONS: We report the use of AI to accurately diagnose blepharoptosis from a clinical photograph with no external reference markers or user input requirement. Most current-generation CNN architectures performed reasonably on this task, with the DenseNet121, and Resnet18 architectures without pre-training performing best in our dataset.
PURPOSE:Blepharoptosis is a known cause of reversible vision loss. Accurate assessment can be difficult, especially amongst non-specialists. Existing automated techniques disrupt clinical workflow by requiring user input, or placement of reference markers. Neural networks are known to be effective in image classification tasks. We aim to develop an algorithm that can accurately identify blepharoptosis from a clinical photo. METHODS: A total of 500 clinical photographs from patients with and without blepharoptosis were sourced from a tertiary ophthalmic center in Taiwan. Images were labeled by two oculoplastic surgeons, with an independent third oculoplastic surgeon to adjudicate disagreements. These images were used to train a series of convolutional neural networks (CNNs) to ascertain the best CNN architecture for this particular task. RESULTS: Of the models that trained on the dataset, most were able to identify ptosis images with reasonable accuracy. We found the best performing model to use the DenseNet121 architecture without pre-training which achieved a sensitivity of 90.1 % with a specificity of 82.4 %, compared to the worst performing model which was used a Resnet34 architecture with pre-training, achieving a sensitivity of 74.1 %, and specificity of 63.6 %. Models with and without pre-training performed similarly (mean accuracy 82.6 % vs. 85.8 % respectively, p = 0.06), though models with pre-training took less time to train (1-minute vs. 16 min, p < 0.01). CONCLUSIONS: We report the use of AI to accurately diagnose blepharoptosis from a clinical photograph with no external reference markers or user input requirement. Most current-generation CNN architectures performed reasonably on this task, with the DenseNet121, and Resnet18 architectures without pre-training performing best in our dataset.