Tae Keun Yoo1, Joon Yul Choi2, Hong Kyu Kim3, Ik Hee Ryu4, Jin Kuk Kim4. 1. Department of Ophthalmology, Aerospace Medical Center, Republic of Korea Air Force, Cheongju, Republic of Korea. Electronic address: eyetaekeunyoo@gmail.com. 2. Epilepsy Center, Neurological Institute, Cleveland Clinic, Cleveland, OH, USA. Electronic address: jychoi717@gmail.com. 3. Department of Ophthalmology, Dankook University Hospital, Dankook University College of Medicine, Cheonan, South Korea. 4. B&VIIT Eye Center, Seoul, South Korea; VISUWORKS, Seoul, South Korea.
Abstract
BACKGROUND AND OBJECTIVE: The purpose of the present study was to investigate low-shot deep learning models applied to conjunctival melanoma detection using a small dataset with ocular surface images. METHODS: A dataset was composed of anonymized images of four classes; conjunctival melanoma (136), nevus or melanosis (93), pterygium (75), and normal conjunctiva (94). Before training involving conventional deep learning models, two generative adversarial networks (GANs) were constructed to augment the training dataset for low-shot learning. The collected data were randomly divided into training (70%), validation (10%), and test (20%) datasets. Moreover, 3D melanoma phantoms were designed to build an external validation set using a smartphone. The GoogleNet, InceptionV3, NASNet, ResNet50, and MobileNetV2 architectures were trained through transfer learning and validated using the test and external validation datasets. RESULTS: The deep learning model demonstrated a significant improvement in the classification accuracy of conjunctival lesions using synthetic images generated by the GAN models. MobileNetV2 with GAN-based augmentation displayed the highest accuracy of 87.5% in the four-class classification and 97.2% in the binary classification for the detection of conjunctival melanoma. It showed an accuracy of 94.0% using 3D melanoma phantom images captured using a smartphone camera. CONCLUSIONS: The present study described a low-shot deep learning model that can detect conjunctival melanomas using ocular surface images. To the best of our knowledge, this study is the first to develop a deep learning model to detect conjunctival melanoma using a digital imaging device such as smartphone camera.
BACKGROUND AND OBJECTIVE: The purpose of the present study was to investigate low-shot deep learning models applied to conjunctival melanoma detection using a small dataset with ocular surface images. METHODS: A dataset was composed of anonymized images of four classes; conjunctival melanoma (136), nevus or melanosis (93), pterygium (75), and normal conjunctiva (94). Before training involving conventional deep learning models, two generative adversarial networks (GANs) were constructed to augment the training dataset for low-shot learning. The collected data were randomly divided into training (70%), validation (10%), and test (20%) datasets. Moreover, 3D melanoma phantoms were designed to build an external validation set using a smartphone. The GoogleNet, InceptionV3, NASNet, ResNet50, and MobileNetV2 architectures were trained through transfer learning and validated using the test and external validation datasets. RESULTS: The deep learning model demonstrated a significant improvement in the classification accuracy of conjunctival lesions using synthetic images generated by the GAN models. MobileNetV2 with GAN-based augmentation displayed the highest accuracy of 87.5% in the four-class classification and 97.2% in the binary classification for the detection of conjunctival melanoma. It showed an accuracy of 94.0% using 3D melanoma phantom images captured using a smartphone camera. CONCLUSIONS: The present study described a low-shot deep learning model that can detect conjunctival melanomas using ocular surface images. To the best of our knowledge, this study is the first to develop a deep learning model to detect conjunctival melanoma using a digital imaging device such as smartphone camera.