Automated detection of diabetic retinopathy using convolutional neural networks on a small dataset
Introduction
Deep learning has enhanced the purpose of computer vision in identifying and classifying images and are a key tool used to automate tasks in our daily lives. Convolutional networks have been consistently developed for object detection, classification, segmentation. The use of convolutional neural networks (CNNs) on medical images has helped the medical sector immensely due to it's ability to learn representations of data [1].
Diabetic Retinopathy turns out to be a major cause of blindness in the western world, and regular screening of the patients reduces the risk of blindness. There are a number of features pertaining to the recognition of retinopathy in fundus photography, and computer vision based trained classifiers work pretty well in classification. Promising work has been displayed in the detection of retinopathy using k-NN classifiers and vector machines. CNNs have also been used for the classification of Diabetic Retinopathy [13,15], given a big dataset and considerable computing power [18]. They have been instrumental in detecting the features such as haemorrhage and hard exudes that identify retinopathy [8]. Deep architectures of CNN have been instrumental in providing the finesse and high performance to trained models by learning patterns from raw images [2]. Due to the availability of annotated data and evolution of GPUs, CNNs have been increasingly applicable in a number of cases. However, in case of medical datasets, huge amounts of annotated data are not readily available yet as the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). Thus, transfer learning has not been very useful for medical datasets as most networks have been trained well to recognize objects present in the ImageNet dataset as shown in Fig. Fig. 1 and Fig. 5, Fig. 6, Fig. 7, Fig. 8, Fig. 9, Fig. 10, Fig. 11, Fig. 12, Fig. 13.
A major problem faced in training the model on less data is underfitting. Moreover, the presence of skewed classes causes the model to overfit on the largest class, which is turn, decreases the the corresponding F1 scores and Cohen's Kappa. Large datasets can often be over-sampled on the lower class, but oversampling on a small dataset will not be of much help against overfitting.
In this paper, we propose a deep learning based CNN method to classify images from a small and skewed dataset of 3050 training images belonging to 4 classes and 419 validation images to achieve a considerably good result. The accuracy metric used by us is Cohen's Kappa.
Section snippets
Related work
There has been considerable work regarding binary classification of Diabetic Retinopathy. Gardner et al used neural networks on 200 images by splitting an image into patches to achieve a sensitivity of 88.4% and specificity of 83.5% for binary classification. His work had been aided by a clinician in classifying the patches prior to using SVM.
Dr.Nayak et al also used neural networks to classify Diabetic Retinopathy Based on 3 classes by recognizing blood vessels and hard exudes from 140 images
Method and structure
We decided upon our network after studying baseline literature [7,22] and testing the performance of other models [[3], [4], [5],9,23]. It was observed that deeper layers cause overfitting as our dataset was comparatively smaller. In our network, we used CNN-based transfer learning on the DenseNet model pre-trained on ImageNet.
Results
The model was validated on 419 fundus images. The validation process was fast. We obtained a validation accuracy of 84.10 %. Accuracy was not used by us as the final metric due to the skewness of data. We chose Cohen's Kappa as our final metric than F1 score as Kappa provides a relativistic accuracy with respect to the other nearby classes, thus imparting a sense a reliability and originality in identification in case of a medical diagnosis [19].
A literature
Discussion and conclusion
Our model has approached the classification of 4-class problem in Diabetic Retinopathy on a small dataset using deep learning. Most earlier algorithms dedicated to the classification of DR fundus images on a small dataset evaded the use of deep learning. Our method has produced comparable results with previous literature given data and hardware constraints and the presence of skewed classes. Transfer learning and fine-tuning on the pre-trained DenseNet has proved to be extremely effective on
Declaration of Competing Interest
None.
References (23)
- et al.
Automated early detection `of diabetic retinopathy
Ophthalmology
(2010) - et al.
Deep learning
Nature
(2015) - Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent...
- et al.
Automated identification of diabetic retinopathy stages using digital fundus images
J. Med. Syst.
(2008) - et al.
Application of higher order spectra for the identification of diabetes retinopathy stages
J. Med. Syst.
(2008) - et al.
Multiclass svm-based automated diagnosis of diabetic retinopathy
- Harry Pratt, Frans Coenen, Deborah M. Broadbent, Simon P. Harding, Yalin Zheng, Convolutional neural networks for...
- Xiaogang Li, Tiantian Pang, Biao Xiong, Weixiang Liu, Ping Liang, Tianfu Wang, Convolutional neural networks based...
- et al.
Algorithms for the automated detection of diabetic retinopathy using digital fundus images: a review
J. Med. Syst.
(2012) - Carson Lam, Darvin Yi, Margaret Guo, Tony Lindsay, Automated detection of diabetic retinopathy using deep learning,...
Cited by (86)
NIMEQ-SACNet: A novel self-attention precision medicine model for vision-threatening diabetic retinopathy using image data
2024, Computers in Biology and MedicineA refined ResNet18 architecture with Swish activation function for Diabetic Retinopathy classification
2024, Biomedical Signal Processing and ControlFrom data to insights
2024, A Biologist's Guide to Artificial Intelligence: Building the foundations of Artificial Intelligence and Machine Learning for Achieving Advancements in Life SciencesDiabetic retinopathy grading review: Current techniques and future directions
2023, Image and Vision ComputingDiabetic retinopathy identification using parallel convolutional neural network based feature extractor and ELM classifier
2023, Expert Systems with ApplicationsA Novel Image Registration Framework for Monitoring Diabetic Retinopathy Progression with Genetic Algorithm-Based Image Alignment
2024, International Journal of Intelligent Systems and Applications in Engineering