An Efficient Transfer Learning Based Approach for Detecting the Abnormal Fundus Images

The vision disabilities are seemingly increasing and prevalent across the world. The current diagnosis require a manual expert for diagnosis. The advancement regarding research in AI for healthcare applications has been focused in recent years. The retinal fundus imaging system is used to capture the retinal images. Those retinal fundus images can be used to detect the vision impairments. Prior research has thoroughly investigated regarding detecting the particular type of disease such as diabetic retinopathy, cataract, glaucoma etc. However, only a little research has been conducted to classify a given retinal image into normal and abnormal fundus image. In this paper, a novel transfer learning based method to detect the abnormal fundus image has been proposed. The EfficientNetV2 has been used as classification model, which aids in classifying the normal and abnormal fundus images. To our knowledge, the EfficientNetV2 has not been used as a transfer learning model before. The proposed model has been evaluated against the recent state-of-the-art models including transformers and multi layer perceptron(MLP) based models which have been found to be working well on the image classification task. It has been observed that convolutional neural networks are still performing better than the recent transformer based models and MLP based models for detecting abnormal fundus images. The proposed model has achieved 95 % accuracy on the dataset received from a hospital.