RetMobileNet: A New Deep Learning Approach for Multi-Class Eye Disease Identification

Human eye is usually called as the window to the soul, it is an intricate organ responsible for the sense of sight [1]. However, its complexity makes it susceptible to a myriad of diseases that can compromise vision [2-4]. From conditions like glaucoma and macular degeneration to diabetic retinopathy [5], the spectrum of eye diseases is vast. The early and accurate identification of these diseases is crucial, as timely intervention can prevent irreversible damage and preserve vision [6-8]. Traditional diagnostic methods, while effective, often rely on the expertise of ophthalmologists and can be time-consuming [9]. Moreover, with the increasing global prevalence of eye diseases, there is a pressing need for more efficient diagnostic tools [10]. The global prevalence of eye diseases such as cataract, glaucoma, and dry eye disease is a major medical, social, and economic problem, especially among the elderly population [11]. The burden of vision loss due to eye diseases has been steadily increasing, especially among those aged 50 years and above [12]. Resource-poor countries bear a disproportionate burden of eye diseases, with blindness due to age-related conditions such as cataract and glaucoma increasing [13]. The global population growth and aging process are contributing to the increasingly severe situation of blindness worldwide [14].

Recent years have witnessed groundbreaking advancements at the intersection of technology and medicine, with deep learning emerging as a transformative tool for medical image analysis [15]. Deep learning, a subset of artificial intelligence, harnesses neural networks to process vast datasets, making it adept at tasks like image recognition [16]. Deep learning is a subset of artificial intelligence (AI) focused on training neural networks to learn from vast amounts of data [17]. These networks consist of layers of interconnected nodes, akin to neurons in the human brain [18]. Through exposure to extensive datasets, deep learning algorithms can autonomously identify patterns and make decisions without explicit programming [19]. This capability renders deep learning particularly effective in tasks such as image and speech recognition, natural language processing, and medical diagnosis [20]. Deep learning is the foundation of CNN, which is a multi-layer network structure simulating the operation mechanism of the biological vision system [21]. Deep learning, with its powerful learning ability to abstract data at a higher level, forms the basis for CNN, allowing it to obtain more accurate data and feature descriptions from original data [22]. CNN is a widely used technique in deep learning for better feature extraction from large datasets, especially in imaging tasks [23].

Convolutional Neural Networks (CNNs) have established themselves as the primary choice for image classification due to their capacity to autonomously and systematically acquire features from images [24]. In the realm of ophthalmology, the application of CNNs to retinal fundus images offers a promising avenue for the automated identification of eye diseases [25]. Retinal fundus images, that capture the back of the eye, including the retina, optic disc, and blood vessels, provide a wealth of information about the health of the eye. Given the intricate details present in these images, CNNs, with their ability to detect subtle patterns, are well-suited for this task [26].

In this study, we conducted a comprehensive assessment of seven prominent CNN models, including VGG19, InceptionV4, Resnet50, MobileNetV1, MobileNetV2, MobileNetV3 Small, and MobileNetV3 Large, to analyze retinal fundus images and determine the most effective model for disease identification. The objective was to determine the most effective model for analyzing retinal fundus images. The results were illuminating. Among the seven models, MobileNetV1 and MobileNetV2 emerged as the superior models, outperforming their counterparts in terms of accuracy and efficiency.

MobileNet architectures, known for their compactness and efficiency, are designed for mobile and embedded vision applications [27]. MobileNetV1 introduced depth-wise separable convolutions that optimize computational efficiency without sacrificing performance [28]. MobileNetV2 further enhances this design with the introduction of the inverted residual structure, ensuring the faster processing times and improved accuracy [29]. Given the stellar performance of MobileNetV1 and MobileNetV2, we developed RetMobileNet, merging their strengths into a powerful tool tailored for multi-class eye disease identification. RetMobileNet 's development is meticulous, incorporating depth-wise convolutional layers for efficient retinal image feature extraction. These layers process each input channel separately to capture intricate details crucial for disease identification. Additionally, we improve the model's performance by implementing batch normalization to stabilize training and speed up convergence. We also introduce pivotal enhancements, such as integrating residual blocks with skip connections. Deep learning models, especially those with numerous layers, can sometimes face the challenge of the vanishing gradient problem, which can impede the model's training [30]. Residual blocks with skip connections address this issue, allowing for more effective training of the deeper layers of the network.

RetMobileNet utilizes five dense layers for final predictions, synthesizing features for accurate identification of eye diseases. The research significantly advances ophthalmology academically by employing deep learning techniques, particularly CNNs, enhancing our understanding of the field. This study thoroughly assesses seven different CNN models to determine their effectiveness in analyzing retinal fundus images, thereby shedding light on the strengths and weaknesses of these models in the context of medical image classification. Furthermore, this research introduces a novel deep learning architecture called RetMobileNet, which combines MobileNetV1 and MobileNetV2 for multi-class eye disease identification. It represents a significant advancement in the design of deep learning models for medical image analysis. Additionally, this study addresses a common issue in deep neural networks: residual blocks with skip connections. By doing this study, it contributes valuable insights to the improvement of deep learning model training. On the practical side, this research has resulted in the development of RetMobileNet, an efficient tool for the automatic identification of eye diseases using retinal fundus images. This streamlines diagnostics has the potential to lessen dependence on human experts. Through optimized feature extraction and improved training, RetMobileNet achieves higher disease identification accuracy, providing substantial advantages for early intervention and vision preservation in real-world healthcare settings.

3.1 Research framework

The research framework (Figure 1) consists of six main stages, namely: Input Fundus Image, Data Preprocessing, Initial Training Phase, Accuracy Optimization, Model Development, and Evaluation and Identification.

1.png

Figure 1. Research framework

In the Input Fundus Image stage, image data was obtained from sources such as public databases. After collecting image data, the next stage involved Data Preprocessing. Preprocessing is a pre-processing process aimed at preparing data before analysis. It involved processes like altering image formats, standardizing pixel intensities, cropping images, or eliminating noise or artifacts from the data. Subsequently, the Initial Training Phase entails the initial training or comparison of seven Convolutional Neural Network (CNN) models (VGG19, InceptionV4, Resnet50, MobileNetV2, MobileNetV1, MobileNetV3 Large, and MobileNetV3 Small) to determine the optimal model. Development or adjustments will then be performed based on the chosen model. After the initial training, the next stage was to optimize accuracy using CLAHE (Contrast Limited Adaptive Histogram Equalization). The use of CLAHE aimed to enhance the contrast and clarity of images with the purpose of improving accuracy in object analysis or recognition. CLAHE enhances local contrast more effectively, making details in various areas of the fundus image more visible. The implementation of CLAHE involves dividing the image into small, non-overlapping areas or tiles, then applying Histogram Equalization to each tile. Contrast is then limited: the histogram is clipped, meaning the maximum value allowed for the histogram bin is set at a predefined threshold to prevent any bin from dominating the histogram, thereby controlling noise amplification. The tiles are then combined using bilinear interpolation to eliminate artificial boundaries between them, resulting in a smooth overall image. By limiting contrast, CLAHE reduces the risk of noise amplification, which is crucial in medical images where clarity and accuracy are paramount. This method facilitates the identification and diagnosis of image conditions by enhancing the visibility of important structures such as blood vessels and the optic disc.

Subsequently, the Model Development Stage was performed that was a further optimization of the selected best CNN model. We found-tunes parameters, such as selecting the optimal learning rate, adding regularization, or applying other optimization techniques. After the development process was complete, we tested the model's performance on a separate testing dataset to obtain an objective evaluation. The final stage was Evaluation and Identification to assess the performance of the created model. Evaluation is conducted to determine the model that can identify or recognize patterns from previously unseen data. By evaluating the model, strengths and weaknesses can be identified, aiding in refining the model for better identification results.

The primary goals of model evaluation involved assessing accuracy, enhancing performance, optimizing the model, and determining its readiness for use. Various evaluation metrics, including accuracy, precision, recall, F1-score, ROC curve, and AUC, can be utilized based on the specific problem type and model goals. Expert verification will be conducted to assess the model's effectiveness in identifying diseases in fundus images.

3.2 Input fundus image

The Fundus Image input is from the MuReD (Multi-Label Retinal Disease) that combines the STARE dataset, RFMiD dataset, and ARIA dataset. The total number of images was 2208, comprising 20 classes [44]. Upon further observation, it was found that there were many missing data, resulting in a total of 1855 available images from the 20 classes. The dataset of fundus images utilized in this research had undergone validation by an eye specialist at SMEC Eye Hospital in Medan, Indonesia.

3.3 Preprocessing data

The goal of Data Preprocessing was to prepare the data for analysis or the creation of Deep Learning models. Proper Data Preprocessing improved the accuracy and reliability of Deep Learning models while preventing errors that may arise from using raw, unprepared data.

3.3.1 Cleaning and normalization

In the previous discussion (Fundus Image Input), after joint observation with fundus image experts, out of the 1855 images available across 20 classes, there were 497 images diagnosed with multiple diseases in one image, and 64 duplicated images. Consequently, these images need to be removed to avoid disrupting the research process. Additionally, out of the 20 classes, we eliminated 8 classes with 94 images due to a very small and imbalanced number of images (as seen in Figure 2).

2.png

Figure 2. Graph of number of images (type of disease)

Based on Figure 2, it can be seen that the top 12 disease classes with the most images include Normal Retina (NORMAL), DN, DR, ODC, Media Haze (MH), Age-Related Macular Degeneration (ARMD), Retinitis (RS), Branch Retinal Vein Occlusion (BRVO), CRVO, Optic Disc Edema (ODE), Central Serous Retinopathy (CSR), and Other Diseases (OTHER). These 12 diseases can be used as the research dataset, while the remaining 8 disease classes with a total of 94 images were not used due to their insufficient quantity. Upon further analysis, it was found that 78 images were blurry and contained light reflections or shadows (lens flare artifacts) during image capture, rendering them unsuitable for the research process as they are considered damaged. After completing the data cleaning process, the dataset was available for research consists of 1122 images.

3.3.2 Image conversion

Based on 1122 images, there were two types of image formats (*.TIFF and *.PNG). Therefore, all images in the *.TIFF format will be converted to *.PNG to ensure uniformity across all images. The conversion from *.TIFF to *.PNG in this research was conducted using the Python programming.

3.3.3 Resize image

Image resizing was carried out due to varying resolutions, resulting in our images being reduced to 224 x 224 pixels. It is done considering the specifications of the research device, that was a notebook with an AMD RYZEN 7 2.49 GHz processor, 8 GB RAM, and a NVIDIA GFORCE GTX 1660Ti 6GB GPU.

3.3.4 Split data

The data split technique consisted of dividing the dataset into three parts, comprising Training, Validation, and Testing data. The Training and Validation data were used as input images during the training process, while the Testing data were separated from the beginning to treat the images used as if they were new data in the identification testing, it is also known as blind testing. The ideal data split selection was performed using the K-Fold Cross-Validation method (K=10) [45]. The Training data split were divided into three parts: 60%, 70%, and 80% of the total image data, while the Validation data were 30%, 20%, and 10%.

3.3.5 Balancing data

The data balancing technique was employed to make the sample count in each group proportional. In this phase, the technique used was oversampling with image augmentation methods on the minority class. The training data after data balancing for each class, utilizing augmentation techniques initially, amounts to 500 images, resulting in a total of 6,000 training data after balancing. The validation data after data balancing for each class, utilizing augmentation techniques, amounts to 50 images. It resulted a total of 600 validation data after balancing.

3.4 Initial training stage

This stage is conducted to assess the initial performance of the CNN model after the data splitting and balancing stages. In the initial training phase, the capabilities of several standard CNN models, namely ResNet50, VGG19, InceptionV4, MobileNetV1, MobileNetV2, MobileNetV3Small, and MobileNetV3Large, will be evaluated using fundus image inputs. This stage resulted comparison of each CNN model's performance.

3.5 Accuracy optimization

The previously trained models based on fundus retina images using 7 CNN architecture models that was optimized for their accuracy using CLAHE. CLAHE, an image processing method was, utilized to improve image contrast while maintaining enhanced details in regions with low contrast levels [46]. It was incorporated as part of the process to enhance accuracy optimization [47]. The CLAHE process consists of three stages, namely: dividing the original image into M×N-sized sub-images, calculating the histogram for each sub-image, and applying clipped histogram on each image. The CLAHE technique involved computing the average number of pixels at each grayscale level, as expressed in Eq. (1) [48], and applying the histogram clip limit, detailed in Eq. (2) [49].

$N_{\text {avg }}=\frac{N_{C R-X p} \times N_{C R-Y p}}{N_{\text {gray }}}$      (1)

$\beta=\frac{M}{N}\left(1+\frac{\alpha}{100}\left(S_{\max }-1\right)\right)$      (2)

where N_avg described the average number of pixels at each grayscale level, N_gray represents the count of gray-level values within the sub-image, N_CR-XP signifies the pixel quantity along the X dimension of the sub-image, and N_CR-YP was the number of pixels in the Y dimension of the sub-image. b was the calculated clip limit, M was the number of pixels in the histogram block, N was the number of intensity intervals (bins) in the histogram, S_max was the maximum possible intensity, and α is a factor influencing the extent to which contrast limitation is permitted.

CLAHE was chosen for optimization due to the following key aspects: Local Contrast Enhancement (Enhances contrast in local areas, crucial for medical imaging like fundus images, making small details such as blood vessels and the optic disc more visible). Noise Control (Limits excessive contrast enhancement, reducing noise amplification and maintaining image clarity). Visual Consistency (Uses bilinear interpolation to combine tiles, producing smoother images without visible boundaries). Discussion of alternative techniques for accuracy optimization: Histogram Equalization (Enhances contrast globally but can cause over-enhancement and excessive noise). Adaptive Histogram Equalization (Enhances local contrast but amplifies noise in homogeneous areas). Gaussian Filtering (Reduces noise but sacrifices sharpness and important details). Median Filtering (Removes impulsive noise but may not preserve sharp edges). Wavelet Transform (Enhances contrast and reduces noise but is complex and computationally intensive). CLAHE is preferred because: Efficiency and Effectiveness (Enhances local details and controls noise efficiently, suitable for precise medical image analysis). Optimal Balance (Balances detail enhancement and noise control without significant added complexity). Proven Performance (Widely tested and used in medical imaging, consistently improving image quality for diagnostics).

Images subjected to CLAHE will undergo retraining. If the outcomes surpass the initial training, they can be compared across different architecture models. If there was no improvement, continuous data augmentation can be employed until superior results were achieved. Improved accuracy will lead to a comparison of the 7 CNN models based on CLAHE-optimized images, identifying MobileNetV1 and MobileNetV2 as the best models for further development.

3.6 Model development (proposed model)

This stage was undertaken to develop the best CNN models previously described, namely, to refine the MobileNetV1 and MobileNetV2 models as the basis for the proposed new model (RetMobileNet). The best CNN models (MobileNetV1 and MobileNetV2) can be enhanced, starting with Depth-Wise Convolution to prevent overfitting and boost computational efficiency while maintaining accuracy. These models undergone the retraining with various parameters like Data Split Ratio, Data Balancing, Image Enhancement, and advanced training hyperparameters (Epoch, Batch_size, Image_size, Weights, Learning rate, Activation Function, Optimizer, Metrics, and Callbacks) until optimal results were achieved. If results remained unsatisfactory, iterations with different parameters and training hyperparameters will be conducted. Additionally, Residual Blocks with Skip Connections will be introduced to help the model learn more concise representations. Further layer modifications, including the addition of 3 Dense Layers and 3 batch normalizations, will aim for an optimal model. It will be followed by the incorporation of an extra Dense Layer and Flatten. Finally, more Dense Layers will be added for making predictions. In total, 5 Dense Layers and 3 Batch Normalizations have been added, resulting in the creation of a new model called RetMobileNet. This model can be reevaluated against the previous 5 CNN models to determine if performance improvements have occurred.

Based on Figure 3, it can be observed that the proposed CNN model, RetMobileNet was constructed from two CNN architectures: MobileNetV1 and MobileNetV2. As a result, RetMobileNet had two inputs during the feature extraction process. The first input in the RetMobileNet model used MobileNetV1, augmented with skip connections to help enhance accuracy and address issues like vanishing or exploding gradients. These skip connections with residual blocks operated by connecting the output from one depthwise separable convolution to the input of the next depthwise separable convolution. In the second input, there was a convolutional layer with 128 filters and a (2,2) stride before being used as input for MobileNetV2. The output from each model included an average pooling layer that was then concatenated using the concatenate technique. It ensured that the features obtained from the outputs of MobileNetV1 and MobileNetV2 were weighted in the fully connected layers, consisting of Dense Layer, batch normalization, and Flatten layer. In the final layer, an additional Dense Layer was added to make predictions (identification) of eye diseases using fundus images.

3.png

Figure 3. RetMobileNet (proposed model)

3.7 Evaluation and identification

The evaluation of the RetMobileNet model in identifying retinal fundus image diseases involved assessing the model's performance in classifying retinal fundus images as normal or identifying the type of disease present. The evaluation was conducted using validation or test data. The evaluation metrics included accuracy (1), precision (2), recall (3), and F1 score (4) [50], to gauge the extent to which the model can accurately recognize and classify retinal fundus images. The way to assess the model's ability to identify images was by conducting tests on a selection of image samples (test data).

Accuracy $=\frac{(A B)-(A C)}{(A B+A D+A C+A E)}$     (3)

Precision $=\frac{A B}{(A B+A E)}$      (4)

Recall $=\frac{A B}{(A B+A D)}$      (5)

$F 1$ Score $=2 * \frac{\text { Recall } * \text { Precision }}{\text { Recall }+ \text { Precision }}$      (6)

where AB: true positive, AC: true negative, AD: false negative, and AE: false positive.

The generalization of RetMobileNet to other datasets and populations may vary, as the obtained results could be specific to the dataset used in this research. Significant advancements in medical imaging, particularly in identifying common eye diseases from retinal fundus images, have been achieved through computational models. The Retina MobileNet (RetMobileNet), which combines the strengths of MobileNetV1 and MobileNetV2, is noteworthy for its transformative potential. With several enhancements such as Depthwise Convolutional layers, Batch Normalization, Residual Blocks with Skip Connections, and the inclusion of 4 Dense Layers, this model has proven to be a robust tool for retina image identification. Its impressive performance metrics, including accuracy, precision, recall, and f1 score, consistently demonstrate scores of over 90%, indicating statistical significance and clinical relevance. The high performance of this model ensures its crucial reliability for early and accurate diagnosis of eye diseases, given their immutable nature. The efficiency and compatibility of the RetMobileNet model make it suitable for real-world clinical settings, especially when rapid and precise diagnosis is paramount. Future research could explore further optimization by integrating RetMobileNet with emerging computational models and implementing real-time feedback mechanisms for continuous refinement based on practical usage. Collaborative studies involving technology developers and ophthalmologists can provide valuable insights to meet evolving clinical needs. Additionally, evaluating the model's application in various medical imaging scenarios can enhance its usability and impact.