Fan Xu1, Yikun Qin2, Wenjing He1, Guangyi Huang1, Jian Lv1, Xinxin Xie1, Chunli Diao1, Fen Tang1, Li Jiang1, Rushi Lan3, Xiaohui Cheng4, Xiaolin Xiao5, Siming Zeng1, Qi Chen1, Ling Cui1, Min Li1, Ningning Tang1. 1. Guangxi Health Commission Key Laboratory of Ophthalmology and Related Systemic Diseases Artificial Intelligence Screening Technology, Ophthalmology Department, the People's Hospital of Guangxi Zhuang Autonomous Region, Nanning, Guangxi, China. 2. China-ASEAN Information Harbor, Nanning, Guangxi, China. 3. Guangxi Colleges and Universities Key Laboratory of Intelligent Processing of Computer Image and Graphics, Guilin University of Electronic Technology, Guilin, Guangxi, China. 4. Guangxi Key Laboratory of Embedded Technology and Intelligent Systems, Guilin University of Technology, Guilin, Guangxi, China. 5. School of Computer Science and Engineering, South China University of Technology, Guangzhou, Guangdong, China.
Abstract
PURPOSE: Infiltration of activated dendritic cells and inflammatory cells in cornea represents an important marker for defining corneal inflammation. Deep transfer learning has presented a promising potential and is gaining more importance in computer assisted diagnosis. This study aimed to develop deep transfer learning models for automatic detection of activated dendritic cells and inflammatory cells using in vivo confocal microscopy images. METHODS: A total of 3453 images was used to train the models. External validation was performed on an independent test set of 558 images. A ground-truth label was assigned to each image by a panel of cornea specialists. We constructed a deep transfer learning network that consisted of a pre-trained network and an adaptation layer. In this work, five pre-trained networks were considered, namely VGG-16, ResNet-101, Inception V3, Xception, and Inception-ResNet V2. The performance of each transfer network was evaluated by calculating the area under the curve (AUC) of receiver operating characteristic, accuracy, sensitivity, specificity, and G mean. RESULTS: The best performance was achieved by Inception-ResNet V2 transfer model. In the validation set, the best transfer system achieved an AUC of 0.9646 (P<0.001) in identifying activated dendritic cells (accuracy, 0.9319; sensitivity, 0.8171; specificity, 0.9517; and G mean, 0.8872), and 0.9901 (P<0.001) in identifying inflammatory cells (accuracy, 0.9767; sensitivity, 0.9174; specificity, 0.9931; and G mean, 0.9545). CONCLUSIONS: The deep transfer learning models provide a completely automated analysis of corneal inflammatory cellular components with high accuracy. The implementation of such models would greatly benefit the management of corneal diseases and reduce workloads for ophthalmologists.
PURPOSE: Infiltration of activated dendritic cells and inflammatory cells in cornea represents an important marker for defining corneal inflammation. Deep transfer learning has presented a promising potential and is gaining more importance in computer assisted diagnosis. This study aimed to develop deep transfer learning models for automatic detection of activated dendritic cells and inflammatory cells using in vivo confocal microscopy images. METHODS: A total of 3453 images was used to train the models. External validation was performed on an independent test set of 558 images. A ground-truth label was assigned to each image by a panel of cornea specialists. We constructed a deep transfer learning network that consisted of a pre-trained network and an adaptation layer. In this work, five pre-trained networks were considered, namely VGG-16, ResNet-101, Inception V3, Xception, and Inception-ResNet V2. The performance of each transfer network was evaluated by calculating the area under the curve (AUC) of receiver operating characteristic, accuracy, sensitivity, specificity, and G mean. RESULTS: The best performance was achieved by Inception-ResNet V2 transfer model. In the validation set, the best transfer system achieved an AUC of 0.9646 (P<0.001) in identifying activated dendritic cells (accuracy, 0.9319; sensitivity, 0.8171; specificity, 0.9517; and G mean, 0.8872), and 0.9901 (P<0.001) in identifying inflammatory cells (accuracy, 0.9767; sensitivity, 0.9174; specificity, 0.9931; and G mean, 0.9545). CONCLUSIONS: The deep transfer learning models provide a completely automated analysis of corneal inflammatory cellular components with high accuracy. The implementation of such models would greatly benefit the management of corneal diseases and reduce workloads for ophthalmologists.
Authors: Travis K Redd; N Venkatesh Prajna; Muthiah Srinivasan; Prajna Lalitha; Tiru Krishnan; Revathi Rajaraman; Anitha Venugopal; Nisha Acharya; Gerami D Seitzman; Thomas M Lietman; Jeremy D Keenan; J Peter Campbell; Xubo Song Journal: Ophthalmol Sci Date: 2022-01-29