Call us Today ! +918886268863 | [email protected]

Automatic Diabetic Retinopathy Diagnosis Using Adaptive Fine-Tuned Convolutional Neural Network

Automatic Diabetic Retinopathy Diagnosis Using Adaptive Fine-Tuned Convolutional Neural Network


Diabetes is one of the main health dilemmas worldwide. One complication of diabetes is diabetic retinopathy (DR), which is one of the main causes of blindness [1]. It has different levels of severity [2] and can be controlled if detected at early stages. DR affects the retina, which is responsible for the conversion of light to the electric signal interpreted to create an image. The retina contains a network of blood vessels that provide nutrition to the retina. Diabetes damages the blood vessels; consequently, the retina does not receive blood supply. This affects the health of the retina and ultimately distorts the eyesight of an individual. The earliest stage of DR is referred to as background retinopathy. At this stage, diabetes does not affect the sight, but impairs the blood vessels. Diabetic retinopathy (DR) is a complication of diabetes that leads to blindness. The manual screening of color fundus images to detect DR at early stages is expensive and time consuming. Deep learning (DL) techniques have been employed for automatic DR screening on fundus images due to their outstanding performance in many applications.


However, training a DL model needs a huge amount of data, which are usually unavailable in the case of DR, and overfitting is unavoidable. Employing a two-stage transfer learning method, we developed herein an intelligent computer-aided system using a pre-trained convolutional neural network (CNN) for automatic DR screening on fundus images. A CNN model learns the domain-specific hierarchy of low- to high-level features. Given this, using the regions of interest (ROIs) of lesions extracted from the annotated fundus images, the first layer of a pre-trained CNN model is re-initialized. The model is then _ne-tuned, such that the low-level layers learn the local structures of the lesion and normal regions. As the fully connected layer (FC) layers encode high-level features, which are global in nature and domain specific, we replace them with a new FC layer based on the principal component analysis PCA and use it in an unsupervised manner to extract discriminate features from the fundus images. This step reduces the model complexity, significantly avoiding the overfitting problem. This step also lets the model adopt the fundus image structures, making it suitable for DR feature detection. Finally, we add a gradient boosting based classification layer.

Existing work:

Employed SURF features, Sreejini and Govindan [3] used BoVW based on K-means, local features (IFT, LBP, and LDP), and color features, and Adal et al. [13] employed Laplacian of Gaussian (LOG) to extract features from fundus images. The hand-engineered features are not directly learned from the data. In addition, their design involves laborious and exhaustive preprocessing and parameter tuning.


As such, this kind of features are not tuned to the lesion structures in fundus images or do not generalize well.

Proposed work:

We propose an effective two-stage method for _ne-tuning a pre-trained CNN model for DR grading using lesion ROIs and fundus images. The method takes a retinal fundus image as input, processes it with the fine-tuned model and grades it into normal or DR levels. The main contributions of the proposed work are as follows.

  • We introduced an intelligent computer-aided diagnosis system based on DL for the DR grading of retinal fundus images, which does not need any preprocessing technique to preprocess the retinal fundus images.
  • We proposed a two-stage fine-tuning method to adapt a pre-trained model to retinal fundus images: in the first stage, it embeds the DR lesion structures in a pre-trained CNN model using lesion ROIs, and in the second stage, it adapts high-level layers to extract the discriminate structures of the retinal fundus images by removing the domain-specific fully connected layer (FC) layers of a pre-trained model and introducing a new PCA layer, which significantly reduces the model complexity and helps overcome the overfitting problem; this also overcomes the limitations of DL model learning due to the small amount of available data for DR detection.

System requirements:

  Software requirements:

  • Operating system           :   Windows.
  • Coding Language :   Python.

Hardware components:

System                 :   Pentium IV 2.4 GHz or intel

Hard Disk                      :   40 GB.

Floppy Drive       :   1.44 Mb.

Mouse                  :   Optical Mouse.

Ram                      :   512 Mb.


We developed herein an automatic system based on deep learning for grading retinal fundus images and referring a DR patient to an ophthalmologist at an early stage. The system was built on a pre-trained model using a two-stage transfer learning method because of the limited available dataset and a huge number of parameters in a deep CNN model. First, we tried three state-of-the-art CNN models pre-trained on ImageNet. Natural images in ImageNet have structures different from those of fundus images; thus, we adapted the hierarchical structure of a pre-trained CNN model to the fundus images by reinitializing the filters of its CONV1 layer using the lesion ROIs extracted from the annotated E-optha dataset and then _ne-tuned it using the ROIs. Second, the FC layers encoded the high-level features relevant to the natural images and have a very large number of learnable parameters.

March 19, 2022

0 responses on "<strong>Automatic Diabetic Retinopathy Diagnosis Using Adaptive Fine-Tuned Convolutional Neural Network</strong>"

Leave a Message

Template Design © VibeThemes. All rights reserved.