Open AccessMulti-Class Retinopathy classification in Fundus Image using Deep Learning Approaches
Author(s) -
Nisha Wankhade,
Kishor K. Bhoyar
Publication year - 2021
Publication title -
international journal of next-generation computing
Language(s) - English
Resource type - Journals
eISSN - 2229-4678
pISSN - 0976-5034
DOI - 10.47164/ijngc.v12i5.454
Subject(s) - computer science , artificial intelligence , convolutional neural network , deep learning , fundus (uterus) , classifier (uml) , machine learning , computation , computational complexity theory , flops , convolution (computer science) , class (philosophy) , contextual image classification , pattern recognition (psychology) , artificial neural network , image (mathematics) , algorithm , parallel computing , medicine , ophthalmology
Retinopathy classification from fundus images put a lot of issues in front of ophthalmologists. Convolution and deep neural network models open the doors to handle such challenges and achieve great success in computer vision, but it is reaching its computational limits. This leads to the rethinking of less computationally intensive network architectures for computer vision problems. In this work we have used a RFMiD dataset, which is challenging for machine learning researchers due its multiclass, multi-labelled, and imbalanced nature. In the proposed work three models are developed to classify the retinopathy from fundus images. The first model inherits the properties of the VGG Net and Inception Net. This results in significant reduction in computational complexity compared with VGG Net and Inception net models. The second model is an improvised version of the previous one with an increase in depth that yields notable improvement in results, while maintaining the lower number of computations. The third model uses a bidirectional LSTM model as a classifier with 192 hand-crafted features. This model gives 0.985 AUC, with a precision of 0.98, and recall of 0.9 respectively.