Enhancing Cataract Detection Precision: A Deep Learning Approach

The process of detecting and grading cataracts from retinal images is a common area of focus in the medical imaging industry and typically follows a three-steps: pre-processing, extraction of relevant features, and final classification as depicted in Figure 4. The first step, preprocessing, enhances the clarity of the fundus images through techniques such as resizing, G-filtering, and normalization. Since fundus images obtained from various sources are of varying sizes, the resizing stage standardizes them to a uniform size, making them suitable for further processing by a CNN. G-filtering involves extracting the G-Channel from the RGB fundus images, which helps eliminate variations in illumination. The G-Channel images retain crucial features of the original images and are clearer than the other channels. Second, the feature extraction step can be performed using either manual or automatic methods. Last, the classification step uses the extracted features to determine the presence and severity of cataracts.

4.jpg

Figure 4. Major steps of cataract identification and classification systems

The literature review is separated into two distinct parts based on the method of feature extraction: (1) manual cataract detection using traditional techniques, and (2) automatic cataract detection using deep learning. In the first section, image processing is utilized to extract hand-crafted features which are then used to classify cataracts through the application of conventional machine learning algorithms. The second section focuses on the use of deep learning algorithms to automatically identify features in fundus images, which are then utilized for training and evaluating machine learning classifiers for cataract classification.

2.1 Cataract detection using conventional method

Conventional methods involve utilizing image processing techniques to extract hand-crafted features and then classifying cataracts using conventional machine learning algorithms. Recently many researchers worked with conventional methods, the work of some of them are discussed below:

Guo et al. [9] used wavelet transform and sketch-based features to develop and evaluate multiclass discriminant analysis algorithms for identifying and categorizing cataracts from fundus images. Their results shows that the accuracy for the wavelet transform-based features is 90.9% for two-class classification and 77.1% for four-class classification, while the accuracy for sketch-based features is 86.1% for two-class and 74.0% for four-class classification.

Cao et al. [10] employed the improved Haar-Wavelet transform to identify suitable features for cataract grading. The approach involved breaking down the four-class classification challenge into three separate two-class classifications through a hierarchical methodology, and subsequently training neural network-based classifiers on each of these two-class problems. The final 4-class classification result is obtained by combining the outcomes from all the 2-class classifiers, resulting in a 2-class accuracy of 94.83% and 4-class accuracy of 85.98%.

Yang et al. [11] developed a method that involved an improved Top-bottom hat transformation to increase contrast and a trilateral filter to decrease noise. The luminance and texture of the images are extracted as features for classification, and a backpropagation neural network (BPN) is employed to identify cataracts using these features. The accuracy of this classification approach is found to be 82.9%.

Yang et al. [12] proposed a cataract detection and grading system that combines multiple machine learning algorithms. The system obtains a wide range of features from fundus images, such as wavelet, sketch, and texture-based features. An ES of a BPN and SVM classifiers are used to make predictions, and the final results are determined through MVS and stacking. The system achieved an accuracy of 93.2% for cataract detection and 84.4% for cataract grading.

Zheng et al. [13] proposed a method for classifying cataracts using 2-D DFT spectrograms of fundus images as classification features. To reduce the dimensionality of the feature vectors, the method employs principal component analysis (PCA) and trains and evaluates the LDA classifier using the Adaboost algorithm. The method achieved a two-class classification accuracy of 95.22% and a four-class classification accuracy of 81.52%.

Fan et al. [14] presented a method utilizing PCA for reducing the feature dimensionality of fundus images obtained through wavelet and sketch-based techniques. The aim of this technique is to minimize computational demands while employing commonly utilized classification algorithms, such as SVMs, Bagging, DT, Gradient Boosting, and Random Forests (RF) to categorize cataracts.

Song et al. [15] introduced a technique for cataract classification that incorporates an advanced semi-supervised learning approach to obtain additional information from unlabeled fundus images of cataracts, along with the standard three image features of texture, wavelets, and sketches. This system integrates multiple binary classifiers into a robust multi-classifier, yielding an accuracy of 88.60% for a 4-class categorization.

Pratap and Kokil [16] investigated a system for cataract detection that employs SVD (singular value decomposition) for feature extraction from fundus images and utilizes a SVM as a classifier. The evaluation of this approach revealed a 2-class classification accuracy of 97.78%.

Nur et al. [17] proposed a technique that leverages the Gray Level Co-occurrence Matrix (GLCM) to extract features of contrast, correlation, energy, and homogeneity from fundus images and uses the k-Nearest Neighbors (k-NN) classifier to identify cataracts. The highest accuracy attained by this method is 80% when k was set to 5.

It can be observed from the preceding discussion that both Slit-lamp and conventional hand-crafted feature-based methods for automatic cataract detection have certain shortcomings:

1. The conventional methods require manual extraction and selection of features, which is a time-consuming process that requires the assistance of ophthalmologists at every stage of cataract detection.

2. The cost of hand-crafted methods is high, and their accuracy is low, limiting their ability to detect only certain types of cataracts.

3. Conventional methods also require a significant amount of time to train the system, as they rely on human help for feature extraction, making it difficult to identify finer details clearly for cataract classifications.

2.2 Cataract detection using deep learning methods

The utilization of deep learning has demonstrated its capability to automatically extract features and overcome limitations in the field of cataract classification and grading. A number of researchers have devoted significant efforts to the development of deep learning models for this purpose, with some of their works being presented in the literature.

Zhang et al. [18] introduced an 8-layer Deep D-CNN to assess its performance and efficiency in automatic cataract detection and classification. The D-CNN consisted of five convolutional layers followed by three fully connected layers, with the output of the last fully connected layer being used as input for a softmax classifier that produces a distribution across four classes. The highest accuracy achieved by the proposed method is 93.52% for cataract detection and 86.69% for cataract grading.

Ran et al. [19] developed an approach for six-level cataract grading using a combination of D-CNN and Random Forests (RF). The DCNN is comprised of three modules for feature extraction at different levels on fundus images, while the RF algorithm performs more precise six-level cataract grading based on the features extracted by the DCNN. The proposed method demonstrates an average accuracy of 90.69% for the six-level cataract grading task.

Li et al. [20] introduced a technique that combined deep learning with ResNet-10 and ResNet-50 to train on 7030 images. The ResNet-50 approach is complex due to its deep layers and large volume of data. The technique achieved 97.25% accuracy for cataract detection through two-level image classification and 87.7% accuracy for cataract grading.

Imran et al. [21] presented a method that uses a combination of Self-Organizing Maps (SOM) and Radial Basis Function (RBF) neural networks for automatic cataract detection and grading. The SOM algorithm is utilized to derive initial centers, which are then fed as inputs to the RBF network for cataract classification and grading. This method achieved 95.3% accuracy for cataract detection and 91.7% accuracy for cataract classification into four classes.

Syarifah et al. [22] improved the performance of the Deep CNN (DCNN) by utilizing a pre-trained CNN architecture known as AlexNet with a lookahead optimizer on SGD and Adam. This method has a two-class classification accuracy of 97.5%. However, the approach faced challenges due to the small size of the dataset used.

Junayed et al. [23] presented CataractNet, a novel convolutional neural network architecture characterized by a reduced number of layers, lower training parameters, and compact kernel sizes. This design strategy aimed to decrease computational time and expense while retaining high accuracy in cataract classification, achieving an accuracy rate of 99.13%.

Pratap and Kokil [24] introduced a method for classifying cataracts into four distinct grades by utilizing transfer learning with a pre-trained AlexNet and an SVM classifier. The approach involved the use of image quality selection parameters to evaluate the quality of fundus images sourced from open datasets. The method demonstrated a high level of accuracy, with a 92.9% success rate in categorizing cataracts into the four classes.

Hasan et al. [25] assessed the capability of four pre-trained CNN models: InceptionResNetV2, Xception, InceptionV3, and DenseNet121, in diagnosing cataracts from retinal images. The results showed that among these models, InceptionResNetV2 exhibited the highest level of accuracy in two-class classification, with a score of 98.17%.

Weni et al. [26] conducted research on a CNN-based approach for cataract detection, with a goal to enhance diagnostic accuracy while mitigating loss. The method achieved a classification accuracy of 95% after conducting 50 epochs.

Simanjuntak et al. [27] carried out an investigation to compare the effectiveness of four CNN architectures, specifically MobileNet, ResNet, GoogLeNet, and a custom CNN model, in classifying cataracts using fundus images. The results indicated that the proposed CNN model exhibited the highest level of accuracy and stability with a score of 92%, followed by MobileNet at 92%, ResNet at 93%, and GoogLeNet at 86%.

Imran et al. [28] presented a hybrid approach combining deep learning models (AlexNet, ResNet, VGGNet with transfer learning) and a SVM classifier for categorizing cataracts into four classes. The proposed architecture achieved an accuracy of 95.65% by utilizing robust feature extraction and classification techniques.

Yadav and Yadav [29] introduced a hybrid system that utilized transfer learning for feature extraction from fundus images. The method employed pre-trained CNN models (AlexNet, ResNet, and VGGNet) and an ES of SVM classifiers to grade cataracts into four categories. The system achieved an accuracy of 96.25% in four-class classification.

It is observed from the above discussion that the efficiency of deep learning approaches in automatic cataract diagnosis and classification has been established as a more efficient alternative to traditional image processing methods. Therefore, this study uses a compact CNN model with an ES of three classifiers namely SVM, NB, DT to classify cataract into four classes (mild, moderate, no, and severe). The advantage of using this architecture is that it can improve the performance of the classification task. The compact representation of CNN model reduces the computation cost by using small kernels, lesser trainable parameters and layers. In addition to this, an ES approach utilized multiple individual classifiers to produce predictions that are combined to attain improved accuracy. SVM, NB, and DT are different types of classifiers that have different strengths and weaknesses, and combining their predictions can help to mitigate these limitations. For example, SVM may be good at handling high dimensional data and non-linear boundaries, while NB may perform better on large datasets. The DT classifier may be better suited for complex relationships and interactions between features. By combining the predictions of these classifiers, the resulting ES approach leverages the strengths of individual classifiers, and produce a more robust result.

In this section, the results and discussions of the proposed method's implementation on fundus image dataset are presented. The experiments are conducted on a personal computer with top-notch specifications including an Intel 7th Generation Core i7-7700 processor, 64-bit Windows 10 operating system, 32GB of RAM and 4GB NVIDIA GTX 1050 Ti graphics card. The proposed method's simulations are carried out utilizing the MATLAB R2019a software, a comprehensive platform that incorporates image processing, deep learning, and machine learning toolboxes.

4.1 Evaluation metrics

In this study, a collected fundus images dataset is used to train and test a method. The performance of the method is then evaluated using several performance metrics: accuracy, precision, sensitivity (or recall), specificity, and F1-score.

$\begin{gathered}{ Accuracy }(A C C)=\frac{T . P .+T . N .}{T . N .+T \cdot P \cdot+F \cdot N .+F . P .} \\ { Sensitivity }(S E N)=\frac{T . P .}{T . P .+F . N .} \\ { Specificity }(S P E)=\frac{T . N .}{T . N .+F . P .} \\ { Precision }(P R E)=\frac{T . P .}{T . P .+F . P .} \\ F 1 {score}(F 1)=2 * \frac{ { Precision } * { Sensitivity }}{ { Precision }+ { Sensitivity }}\end{gathered}$

Accuracy is a metric that quantifies the number of accurate predictions made by a model. Precision assesses the ratio of true positive predictions to all positive predictions made by the model, while sensitivity (recall) evaluates the proportion of actual positive instances that were correctly identified by the model. In the medical field, precision and recall are deemed as the two crucial evaluation metrics. The F1-Score combines precision and recall through the harmonic mean and represents the balance between precision and recall. A F1-Score that is high signifies a good balance between precision and recall. In the field of medical diagnosis, avoiding false negatives (incorrectly reporting that a patient does not have a condition when they actually do) is of utmost importance, thus recall is prioritized over precision (correctly identifying patients who have the condition). False negatives could result in serious consequences for the patient, whereas false positives (incorrectly reporting that a patient has a condition when they do not) can be resolved through additional testing. Therefore, this study takes into account a comprehensive set of performance evaluation metrics, including accuracy, precision, recall, specificity, F1-score, and AUC, in order to assess the effectiveness of the diagnostic process.

4.2 Performance evaluation

In this study, the efficiency of the proposed methodology is evaluated by comparing its feature extraction and classification capabilities. It is observed that, as there are no benchmark datasets available for comparison, all the algorithms are trained and evaluated using a self-collected dataset.

In this study, the performance of the proposed method is evaluated based on feature extraction using pre-trained CNN architectures such as AlexNet-Softmax, VGGNet-Softmax, ResNet-softmax, AlexNet-SVM, VGGNet-SVM, and ResNet-SVM. The results, displayed in Table 4, indicate that the proposed method outperformed the other models evaluated. The proposed method demonstrated exceptional results, with a high accuracy score of 97.34%, sensitivity of 95.63%, specificity of 97.92%, precision of 93.87%, and F1-score of 94.74%. These results show that the proposed method exhibits remarkable performance compared to other algorithms.

Table 4. Performance comparison between the proposed methodology and pre-existing CNN models

This study also evaluated the performance of the proposed method in terms of classification by using a custom CNN architecture and different classifiers, including SVM, NB, and DT. The results, as displayed in Table 5, demonstrate that the proposed method surpasses the performance of the other classifiers.

The reason for the improved performance is due to the utilization of various classifiers, each having its unique strengths and weaknesses, resulting in a more comprehensive and diverse representation of the underlying data. Therefore, an ES of classifiers averaged out the errors made by individual classifier. Thus, proposed method used an ES of three classifiers (SVM, NB, DT) that combined the predictions of all classifiers using MVS to make a more accurate prediction to grade the cataracts.

Table 5. Performance comparison of the proposed methodology using various classifiers

4.3 Computation time

The computation time for each stage of the proposed methodology is documented in Table 6. The findings reveal that the feature extraction process requires a duration of 0.35 seconds, while the classification of a single retinal test image takes 0.60 seconds, resulting in a total computation time of 0.95 seconds. The comparison chart depicted in Figure 10 illustrates the superiority of the proposed method in terms of speed, as compared to previous approaches. This is achieved through the implementation of compact kernels, a reduced number of training parameters, and fewer layers, in CNN which facilitate the rapid extraction of features from retinal images.

Table 6. The computation time for each step in the proposed methodology

4.4 Result analysis and discussion

This method initially uses preprocessing and augmentation techniques to elevate the image quality and enlarge the size of the dataset, subsequently selecting 800 high-quality images per class in consultation with an expert ophthalmologist. The augmented dataset is subsequently divided into 80:20 for training and testing data, respectively. During the training phase, the CNN model is trained on the training data in batches of 32 and the network weights are optimized within the range of [-1, 1]. The learning rate is set at 3×10^-4 and dropout values are established as 0.4, 0.4, 0.5 to deter overfitting. The other hyper-parameters are established as indicated in Table 7. For cataract classification, an ES of NB and DT, including linear kernel SVM classifier for CNN model. The linear kernel function is selected as it performed well with a large number of features and reduced computational cost. During the testing phase, the final outcome is established through the MVS approach based on the predictions of all classifiers. The results indicate that the proposed method exhibits superiority over all pretrained CNN models with softmax classifier by a margin of approximately 5 to 6%, and all pretrained CNN models with SVM classifier by a margin of approximately 2 to 3%, as well as custom CNN models with individual classifiers in the ES by a margin of 1 to 3% in terms of accuracy. The computation time encompasses both feature extraction and image classification time. Figure 10 presents a comparison of the computation time for various methods. It can also be observed from Table 6 and Figure 10 that the time required to classify a test sample using a AlexNet-softmax classifier, VGG19-softmax classifier, ResNet50- softmax classifier, AlexNet-SVM classifier, VGG19- SVM classifier, ResNet50- SVM classifier, and the proposed method is 1.15 seconds, 1.25 seconds, 1.35 seconds, 1.20 seconds, 1.30 seconds, 1.40 seconds and 0.95 seconds respectively. This indicates that the classification of fundus images with the proposed method is faster than the other methods. The reason of less computation time of the proposed method compare to other pre-trained networks is the use of the custom CNN network that has lesser layers, small size kernels and fewer trainable parameters.

10.png

Figure 10. Comparative chart of computation time (s) of proposed method with other methods (one image)

Table 7. Fine-tuned hyperparameters details

To evaluate the efficacy of the proposed method, its performance is compared with state-of-the-art techniques. Table 8 provides a comparison of different cataract classification methods using common performance metrics. Imran et al. [42] created an automated system for identifying various levels of cataracts by utilizing a D-CNN in conjunction with data augmentation methods. The standard data augmentation and Gaussian scale space theory methods are utilized to improve the quantity and quality of the data. This approach achieved an accuracy of 93.79%. Another method by Cao et al. [10] used an enhanced Haar wavelet transform to extract features and achieved an accuracy of 85.98% in four-class cataract grading. Pratap and Kokil [24] invented a method that performs image quality assessment, then select only good-quality images for further processing. A pre-trained AlexNet and an SVM classifier are utilized for feature extraction and classification, achieving an accuracy of 92.87%. Imran et al. [28] used a hybrid model that combines three pre-trained CNNs (Alexnet, VGGNet, ResNet) and one SVM classifier, resulting in 95.63% accuracy. Yadav and Yadav [29] presented hybrid approach, leveraging various pre-trained CNNs (AlexNet, VGGNet, ResNet) with transfer learning for feature extraction. The feature vectors generated by each network individually, as well as in a fused form, are applied to ES of SVM classifiers for 4-stage cataract classification. The accuracy obtained by this method was 96. 25%. The data presented in Table 8 demonstrates the superiority of the proposed methodology in comparison to other state of art methods. The evaluation of the proposed methodology can be visualized through the confusion matrix depicted in Figure 11(a) and ROC displayed in Figure 11(b).