Eye Disease Net: an algorithmic model for rapid diagnosis of diseases

ED_Resnet module

The original ResNet (Wightman, Touvron & Jégou, 2021) residual residual convolution module and our ED_Resnet convolution module are shown in Fig. 4, respectively. We take inspiration from the figure in (a) and design a new module (b). Our module consists of three convolutions, a 1 × 1, a 5 × 5 depth convolution and a 1 × 1 convolution, while the input image values are directly concatenated with the output of the feature map after three convolutions by jumping to understand the output.

ED_Xception module

As shown in Fig. 5, we refer to the Xception (Chollet, 2017) network architecture, as shown in Fig. 5A, and experimentally optimize Xception’s backbone by using 5 × 5 and 7 × 7 convolutions instead of the original 3 × 3 convolutions. Although there is a slight increase in computational effort, we finally adopt the network structure diagram in Fig. 5B during the experiment.

Activation functions

In order to further ensure the stability of the model, we refer to the characteristics of the activation functions of ReLU (Chen et al., 2020) and ReLU6 (Mansuri et al., 2022), optimize on the swish (Sudharsan & Ganesh, 2022) activation function to obtain a new activation function swish6, and apply swish6 to the ophthalmic image classification task, and obtain better classification results. It is well known that the swish activation function formula is shown in Eq. (2).

(2) $$F\left(x\right)=x\cdot Sigmoid\left(\beta x\right)$$where x represents the input pixel value and β is a constant or trainable parameter. The swish activation function is shown in Fig. 6, with separate plots of the swish activation function for values of β of 0.1, 1.0 and 10.0.

We can see from Fig. 6 that as the input value increases, the output value increases. However, when the input value exceeds a certain value, there is a partial gradient explosion in the output value. In view of this, we set a threshold in the swish activation function so that the output value y is fixed at 6 when the input value is greater than 6. The thinking comes from ReLU6 as an improvement on ReLU. Finally, our activation function swish6 is shown in Eq. (3).

(3) $\text{y}(\text{x})=\begin{array}{ll}x\cdot sigmoid(\beta x)& \text{x}\hspace{0.17em}<6\\ 6& \text{x}>\text{6}\end{array}$

Structure of the ED_Xception model

In order to provide a better description of the overall structure of our model, we have produced Table 1 and interpreted the ED-Net network model in detail by Layer, Intput Size, Output Size, Kernel Size, Stride and Padding. As can be seen from Table 1, the structure of our ED-Net network model consists of one ED_Conv, two ED_Resnet, three ED_Xception and one ED-Linear. The input image is taken from a 3 × 3 feature map and the ED-Net network model eventually outputs a seven-class output probability.

To further visualise the structure of the ED-Net network model more intuitively, we visualised the ED-Net output as shown in Fig. 7. We have substituted different colours for different models, while the input value is a 224 × 224 sized, three-channel image, and the final output value of each module is annotated in the figure as the input value for the next module.

Datasets

Our experimental dataset comes from the public dataset of the Kaggle Competition (https://www.kaggle.com/datasets/kondwani/eye-disease-dataset; https://www.kaggle.com/datasets/gunavenkatdoddi/eye-diseases-classification) and Shenzhen Aier Eye Hospital Affiliated to Jinan University, our dataset is divided into a total of seven categories, namely: Bulging_Eyes, There were 4,600 images in our dataset, 3,451 in the training set and 1,149 in the test set, and the details are shown in Table 2. The horizontal and vertical resolutions of the images are 96 dpi and the bit depth is 24 bits.

Experimental environment

We used Ubuntu 20.04.4 LTS 64-bit operating system and NVIDIA 3080Ti 12G video memory for our experiments. We used an input size of 224 × 224 per image, a batch size of 64 for each input model, a uniform algorithm optimizer set to SGD (stochastic gradient descent) (Woodworth et al., 2020), an initial learning rate of 0.01, and a learning rate of 95% of the original for each iteration, an epoch of 100 for the number of iterations, a momentum of 0.9, and a weight decay of 0. 100, momentum was set to 0.9, weight decay was set to 0.0004, and the loss function was Cross Entropy Loss Function (CELF) (Ho & Wookey, 2019).

Comparison of experimental data

To further verify the validity of this article, we produced Tables 3–6. we compared the algorithms with classical ones (Vgg16 (Sitaula & Hossain, 2021), Resnet50 (Wen, Li & Gao, 2020), Densenet121 (Zhang et al., 2019), ResNext_34×4d-50 (Zhou, Zhao & Wu, 2021), ShuffleNetV2 (Vu, Le & Wang, 2022) and Mobilenetv3_ large (Chen et al., 2022)), advanced algorithms (Conformer (Guo et al., 2021), RepMLP_B224 (Ding et al., 2021a), RepVGG_D2se (Ding et al., 2021b), ConvMixer (Ng et al., 2022), and Hornet-L-GF (Rao et al., 2022)), transformer algorithms (DeiT-base (Touvron et al., 2021), PoolFormer_M48 (Yu et al., 2022), SVT_large (Fan et al., 2022), EfficientFormer-l7 (Li et al., 2022b) and MViTv2_large (Li et al., 2022a)) and similar algorithms for experimental comparison.

As can be seen from Table 3, our algorithm ED-Net is more accurate than all the classical algorithms, while the floating point operations (FLOPs) are all lower than the others except for being on par with ShuffleNetV2. In terms of Param (parameter), our algorithm is 51.6 times lower than Vgg16 and on par with ShuffleNetV2. This indicates that our algorithm is more suitable for application to the ophthalmic disease classification task than the classical algorithm.

As can be seen from Table 4, ours has a high accuracy rate compared to advanced CNN algorithms in terms of accuracy, although our accuracy rate is lower compared to the Hornet-L-GF algorithm and equal to ours compared to the RepVGG_D2se algorithm. Also, on FLOPs, our algorithm ED-Net is 216.1 times less accurate compared to Hornet-L-GF and 228.5 times less accurate compared to RepVGG_D2se; on Param, our algorithm is 73.2 times less accurate compared to Hornet-L-GF and 49.8 times less accurate compared to RepVGG_D2se. Thus, we further demonstrate that our algorithm ED-Net is more suitable for ophthalmic disease classification tasks than advanced CNN algorithms.

It is well known that the transformer algorithm has obtained many superior results in numerous fields, therefore, we compared the transformer algorithm in the course of the algorithm. As can be seen from Table 5, the transformer algorithm performed better overall on the ophthalmic disease classification task, with the highest accuracy being the MViTv2_large algorithm, but the high FLOPs and Param of the MViTv2_large algorithm indicate that the complexity of the algorithm is high and is not suitable for application to the ophthalmic disease classification task. deiT-base, PoolFormer_M48, SVT_large and EfficientFormer-l7 algorithms are all lower in accuracy than our algorithm ED-Net, while our algorithm is still significantly lower than the other transformer algorithms in terms of FLOPs and Param. Therefore, our algorithm ED-Net is more suitable for ophthalmic disease classification tasks than the transformer algorithm.

To further validate the effectiveness of our algorithm, we further conducted experimental comparisons with similar algorithms for ophthalmic diseases, while our experimental code was taken from the two best performing algorithms in the kaggle competition, code from literature (Yousrry, 2022) and literature (Fakher, 2021), respectively. We can see from Table 6 that our algorithm ED-Net still performs the best in terms of accuracy, 4.46% higher compared to Yousrry (2022) and 3.54% higher compared to literature (Fakher, 2021), while our algorithm significantly outperforms literature (Yousrry, 2022) and is on par with literature (Fakher, 2021) in terms of FLOPs and Param. Therefore, we further validate that our algorithm ED-Net has a better performance and is suitable for application in ophthalmic disease classification tasks.

Visualisation effects

To further verify the validity of the data of the algorithm in this article, we visualized the process of ED-Net’s accuracy and loss change. It can be seen from Fig. 8 that the ED-Net network model is stable in the network region at 100 epochs of iterations, which proves that we set a good number of iterations. In addition, as the number of iterations keeps increasing, the accuracy of the ED-Net network model keeps increasing and the loss keeps decreasing, therefore, the ED-Net network training process is normal, which indicates that our ED-Net network model is set up more reasonably. Therefore, the ED-Net network model we set up was reasonable in the training and testing phases, further validating the reliability of our data.