EfficientNet-B0-Based Deep Learning Framework for Four-Class Cataract Severity Classification Using Fundus Images

One of the most common causes of impairment of vision and blindness all over the world is cataract [1]. This develops gradually as the lens becomes cloudy and the quality of vision and activities in the normal course of daily living decrease [2]. A trained ophthalmologist is typically necessary for clinical assessment, and specialist ophthalmic service may be less readily available in rural and resource-poor areas [3]. The difficulties have spurred computer-aided cataract screening and cataract severity assessment [4]. Figure 1 illustrates the global burden of vision impairment and cataract, based on estimates reported by the World Health Organization [5].

Fundus photography is widely used in ophthalmic screening and provides an accessible image-based modality for automated analysis. Although cataract is primarily associated with lens opacity, it can influence the appearance and quality of retinal photographs. Consequently, previous studies have investigated fundus images for automated cataract detection and severity grading. Global estimates for adults aged 50 years and older identify cataract as a major contributor to blindness, as shown in Figure 2 [6]. In India, untreated cataract was reported as the leading cause of blindness among adults aged 50 years and older, accounting for 66.2% of blindness cases in the National Blindness and Visual Impairment Survey 2015–2019; the distribution of major causes is shown in Figure 3 [7].

The Ocular Disease Intelligent Recognition 5K (ODIR-5K) dataset comprises 10,000 color fundus photographs from 5,000 patients, including images from both eyes and diagnostic information for multiple ocular conditions [8]. Class imbalance is an important challenge in multi-class medical image classification because overall accuracy may not adequately reflect minority-class performance. Therefore, the present study uses class-wise and macro-averaged measures alongside overall accuracy and interprets severe-class performance cautiously because of its limited representation in the independent test set.

Researchers have investigated fundus photography for automated cataract assessment using conventional machine learning (ML) and deep learning (DL) approaches [9-13]. Convolutional Neural Network (CNN)-based and transfer learning (TL) methods have further demonstrated the feasibility of cataract detection and severity grading from ophthalmic images [14-17]. However, differences in datasets, severity definitions, class distributions, model architectures, and evaluation protocols limit direct comparison between reported results. Reliable evaluation of minority severe-cataract cases also remains challenging.

EfficientNet-B0 provides a balance between classification performance and computational requirements through compound scaling of network depth, width, and input resolution [18]. Accordingly, this study investigates EfficientNet-B0 for four-class cataract severity classification from fundus images. Three configurations—Frozen Baseline EfficientNet-B0, Severe-Focused EfficientNet-B0, and Final Frozen EfficientNet-B0—are evaluated using the same independent test observations. Their performance is examined using overall, macro, weighted, class-wise, agreement-based, statistical, and computational analyses. This research highlights an EfficientNet-B0-based approach to classify fundus images into four classes: Normal, Mild, Moderate, and Severe cataract.

The study focuses on independent-test and minority-class analysis, with particular attention to the underrepresented Severe category. Statistical and computational evaluation includes paired statistical testing and measurements of parameter count, FLOPs, model size, inference time, and throughput. Finally, the Final Frozen EfficientNet-B0 configuration is selected based on its overall predictive performance, class-wise behavior, and computational characteristics while explicitly recognizing the uncertainty associated with the limited number of Severe cases.

In Section 2, the existing literature in the area of cataract analysis is reviewed, and the research gap to be addressed by this work is set. Section 3 explains the methodological approach, which includes data preparation, model configurations, network design, training methods, and performance evaluation. The experimental results are presented in Section 4, with an analysis of the results of the various EfficientNet-B0 architectures. The paper ends with a summary of the main findings, a discussion of the contribution and limitation(s) of the study, and suggested directions for future studies in Section 5.

The classification task includes Normal, Mild, Moderate, and Severe categories. Three configurations—Frozen Baseline EfficientNet-B0, Severe-Focused EfficientNet-B0, and Final Frozen EfficientNet-B0—are evaluated under a common experimental protocol using the same independent test set. Training and validation data are used for model development and parameter selection, while the independent test set is reserved for final evaluation to prevent test information from influencing model development. Performance is assessed through predictive, class-wise, statistical, and computational analyses.

3.1 Dataset and class distribution

The present study uses a curated ODIR-5K-derived fundus-image dataset for four-class cataract severity classification. The final dataset contains 3,868 labelled fundus images from 2,029 unique patients. The four-class labels were established during dataset curation based on the available diagnostic information and predefined categorization criteria used for the experimental dataset. The labels were used as experimental categories and were not independently established through a separate clinical grading procedure or ophthalmologist validation. Duplicate, invalid, and unusable samples were removed during dataset preparation.

Let the complete dataset be represented as:

$D=\left\{\left(x_i, y_i\right)\right\}_{i=1}^N$     (1)

where, $x_i$ denotes the i-th fundus image, $y_i$ denotes its corresponding severity class, and $N=3868$. The resulting dataset comprised 3,466 Normal, 195 Mild, 179 Moderate, and 28 Severe images, indicating substantial class imbalance. A patient-independent partitioning strategy was adopted, ensuring that images from the same patient were not assigned to different subsets. The final partition included 2,698 training images from 1,437 patients, 585 validation images from 296 patients, and 585 independent test images from 296 patients, representing 2,029 unique patients overall, with no patient overlap among the subsets. The training and validation sets were used for model development, while the independent test set served as a common evaluation set for all three EfficientNet-B0 configurations. The proposed EfficientNet-B0-based framework for four-class cataract severity classification is shown in Figure 4. Table 2 summarizes the distribution of images across the training, validation, and independent test subsets.

Figure 4. Proposed experimental framework for cataract severity classification

Table 2. Dataset overview and class distribution

No resampling or augmentation was applied to the validation or independent test subsets. The independent test set contained 511 Normal, 31 Mild, 38 Moderate, and 5 Severe images. Owing to the limited representation of the Severe category, class-wise and macro-averaged metrics were evaluated alongside overall accuracy, and the Severe-class results were interpreted cautiously. Representative fundus images from the four severity categories are presented in Figure 5.

Figure 5. Fundus images: (a) Normal, (b) Mild, (c) Moderate, (d) Severe

3.2 Image quality assessment and selection

To further characterize the quality of the RGB fundus images, Natural Image Quality Evaluator (NIQE) and Perception-based Image Quality Evaluator (PIQE) were used as no-reference image quality measures. In general, lower values of these metrics correspond to comparatively better perceptual image quality. For the quality assessment performed in this study, images satisfying both NIQE and PIQE were categorized within the selected acceptable-quality range, while samples not satisfying these criteria were treated as comparatively lower-quality images. The resulting quality-score distributions and selection criteria are illustrated in Figure 6, with Figure 6(a) presenting the threshold-based assessment and Figure 6(b) showing the corresponding NIQE–PIQE scatter distribution.

(a)

(b)

Figure 6. Resulting quality-score distributions and selection criteria: (a) threshold-based fundus image quality selection, (b) scatter plot showing the quality distribution of fundus images

As illustrated in Figure 7, region P represents the selected acceptable quality region in the NIQE–PIQE space. Images with coordinate points falling within region P satisfy the defined quality criteria, whereas those outside the region are considered comparatively poor-quality samples. This NIQE–PIQE analysis was used solely as an additional image quality assessment and visualization step and did not alter the patient-independent data partitioning, model training procedure, or the reported evaluation results of the Final Frozen EfficientNet-B0 configuration.

Figure 7. Image quality assessment analysis of the dataset

3.3 Image preprocessing

The fundus images were prepared using a consistent preprocessing pipeline before being provided to the EfficientNet-B0 model. All images were resized to 224 × 224 pixels to match the input dimensions of the backbone network and were converted into numerical tensors.

$x_i^r=R\left(I_i, 224,224,3\right)$     (2)

where, $I_i$ is the original fundus image, and $R(.)$ denotes the resizing operation.

Each input image is therefore represented as an RGB image with three color channels. The RGB representation and its individual R-, G-, and B-channel components are illustrated in Figure 8.

Figure 8. RGB components of a fundus image

Each input image is therefore represented as:

$x_i^r \in R^{224 \times 224 \times 3}$

The images were then processed using the preprocessing function associated with EfficientNet-B0. The four severity categories were encoded as integer labels representing Normal, Mild, Moderate, and Severe.

During model development, data augmentation and controlled minority-class sampling were applied only to the training data to reduce the effect of class imbalance. Training data were shuffled and batched, while validation and test data were used in their original patient-independent form. Prefetching was applied to improve data-delivery efficiency during training and evaluation.

$\tilde{x}_{\imath}=T\left(x_i ; \theta_i\right)$     (3)

where, $T(.)$ represents the augmentation transformation, and $\theta_i$ denotes the transformation parameters. $T$ includes the augmentation operations actually used in the final experiment, while validation and test images are not augmented.

Following image resizing and training-only augmentation, the preprocessed images were passed through the preprocessing function associated with EfficientNet-B0. Let $x_p^i$ denote the resized and augmented image. The EfficientNet-B0 preprocessing operation can be represented as

$x_i^e=P_{\text {EffNet }}\left(x_p^i\right)$     (4)

where, $P_{EffNet}(.)$ denotes the preprocessing function implemented for the EfficientNet-B0 input pipeline and $x_i^e$ represents the resulting model-ready image.

This preprocessing pipeline provided a consistent input representation for all three EfficientNet-B0 configurations.

The validation and independent test sets remained unchanged throughout model development and final evaluation, ensuring that augmentation and minority-class sampling did not influence the reported test performance.

3.4 Proposed Final Frozen EfficientNet-B0 architecture

EfficientNet-B0 is used as the backbone network for the proposed four-class cataract severity classification framework. The architecture was selected because it provides an effective balance between feature representation capability and computational requirements through compound scaling of network depth, width, and input resolution. In the present study, the ImageNet-pretrained EfficientNet-B0 backbone is adapted to classify fundus images into four severity categories: Normal, Mild, Moderate, and Severe. The overall architecture of the proposed Final Frozen EfficientNet-B0 framework is illustrated in Figure 9.

Figure 9. Overall architecture of the proposed Final Frozen EfficientNet-B0 framework

The input to the network is a preprocessed fundus image of size. The EfficientNet-B0 backbone progressively extracts hierarchical visual features from the input image. The feature extraction process can be represented as:

$F=f_\theta(X)$     (5)

where, $X$ represents the preprocessed fundus image, where $f_\theta$ denotes the EfficientNet-B0 feature extractor, and $\theta$ represents its learnable parameters, and $F$ denotes the extracted feature maps.

The feature maps generated by the backbone are subsequently aggregated using Global Average Pooling (GAP), which summarizes the spatial information in each channel into a single feature value and produces a compact feature representation. For the k-th feature channel, this operation is expressed as:

$g_k=\frac{1}{H W} \sum_{i=1}^H \sum_{j=1}^w F_k(i, j)$     (6)

where, $H$ and $W$ denote the height and width of the feature map, $F_k(i, j)$ represents the activation at location ( $i, j$ ) within the $k$-th channel and $g_k$ is the resulting pooled feature value. By averaging the activations over the spatial dimensions, GAP produces a lower-dimensional representation while preserving the information encoded within each feature channel.

The resulting pooled features are subsequently passed through a dropout layer with a rate of 0.30, which is used during training to improve regularization and reduce the likelihood of overfitting.

The dropout operation can be represented as:

$\tilde{g}=m \odot g$     (7)

where, $g$ represents the pooled feature vector, $m$ denotes the dropout mask, ⊙ represents element-wise multiplication, and $\tilde{g}$ is the regularized feature vector.

The resulting feature vector is subsequently passed to the final four-class classification layer.

A fundus image is resized to 224 × 224 × 3, followed by data augmentation during training and EfficientNet-specific preprocessing. The preprocessed image is then passed through the EfficientNet-B0 backbone, consisting of an initial convolutional stem and multiple MBConv blocks for hierarchical feature extraction. The extracted features are processed using Global Average Pooling and dropout before being passed to a fully connected layer with four output units. A softmax activation function produces the predicted probabilities for the Normal, Mild, Moderate, and Severe categories.

3.5 Four-class classification

The final classification layer maps the regularized feature representation to the four cataract severity categories: Normal, Mild, Moderate, and Severe. The output logits for an input image are expressed as

$z=W \tilde{g}+b$     (8)

where, $\widetilde{g}$ represents the regularized feature vector after dropout, where $W$ and $b$ correspond to the learnable weight matrix and bias vector of the final fully connected layer, respectively, and $z$ denotes the resulting vector of class logits. These logits are subsequently transformed into class probabilities using the softmax function:

The probability associated with the c-th class is computed by applying the softmax transformation to the output logits:

$P(y c \mid X)=\frac{e^{z_c}}{\sum_{j=0}^3 e^{z_j}}, c \in\{0,1,2,3\}$     (9)

The model assigns the input image to the class with the highest predicted probability among the Normal, Mild, Moderate, and Severe categories:

$\hat{y}=\arg _c^{\max } P(y=c \mid X)$     (10)

3.6 Loss function and training strategy

3.6.1 Loss function

Since the classification task contains four mutually exclusive classes and integer class labels are used, sparse categorical cross-entropy is employed as the training loss.

The loss for N training samples is defined as

$L_{C E}=-\frac{1}{N} \sum_{i=1}^N \log \left(\hat{p}_{i, y_i}\right)$     (11)

where, $y_i$ is the ground-truth class label of the $i$-th image and $\hat{p}_{i, y_i}$ is the predicted probability assigned to its class. Optimization of this loss function promotes higher predicted probabilities for the correct cataract severity category.

3.6.2 Model optimization

The trainable model parameters are optimized by minimizing the classification loss:

$\theta^*=a r g_{\min _\theta L_{C E}(\theta)}$     (12)

where, $\theta^*$ represents the optimized model parameters.

The models were optimized using sparse categorical cross-entropy as the objective function and the Adam optimizer. The optimization process was monitored using validation performance, with early stopping and adaptive learning-rate adjustment employed to support stable convergence and reduce unnecessary training.

3.6.3 Computational environment

Experiments were implemented using Python 3.12.13, TensorFlow 2.20.0, Keras, and scikit-learn, and executed in Google Colab using an NVIDIA Tesla T4 GPU.

3.6.4 Training and freezing strategy

The three EfficientNet-B0 configurations were developed using the same basic optimization framework but differed in their training and parameter-freezing strategies. Frozen Baseline EfficientNet-B0 served as the reference configuration, while Severe-Focused EfficientNet-B0 incorporated a training strategy intended to improve recognition of the underrepresented severe class. Final Frozen EfficientNet-B0 used the selected layer-freezing strategy, in which designated layers remained non-trainable while the remaining parameters were updated during optimization.

The three configurations were subsequently evaluated using the same patient-independent test set to examine the effect of the different training strategies on four-class cataract severity classification.

3.6.5 Training configuration and hyperparameters

All experiments used an EfficientNet-B0 backbone initialized with ImageNet-pretrained weights. Each input fundus image was resized to 224 × 224 × 3 and processed using the preprocessing procedure associated with EfficientNet-B0. Training was performed with a learning rate of 1 × 10⁻⁵, a batch size of 32, and a maximum of eight epochs. Early stopping was configured with a patience of three epochs, while the learning rate was reduced by a factor of 0.2 after two consecutive epochs without improvement in the monitored validation performance. The three experimental configurations followed this common training protocol and differed primarily in their training-data sampling procedures and backbone layer-freezing strategies.

3.7 Evaluation metrics

Because the dataset is highly imbalanced, particularly for the severe category, model performance is evaluated using multiple complementary metrics rather than overall accuracy alone.

3.7.1 Accuracy

Overall accuracy represents the proportion of test observations that are correctly classified by the model:

Accuracy = $\frac{N_{\text {correct}}}{N_{\text {total}}}$     (13)

For the four-class problem, $N_{\text {correct}}$ represents the number of test images for which the predicted class matches the ground-truth class.

Although accuracy provides an overall measure of classification performance, it can be dominated by the majority normal class in an imbalanced dataset. Therefore, additional class-balanced measures are considered.

3.7.2 Precision

Precision quantifies the reliability of predictions assigned to a given class. Specifically, for class c, it is defined as the fraction of correctly classified samples among all samples predicted as belonging to that class.

For class, precision is defined as:

Precision$_c=\frac{T P_c}{\left(T P_c+F P_c\right)}$     (14)

where, $T P_c$ indicates the number of correctly identified samples for class $c$ and $F P_c$ denotes the number of samples incorrectly predicted as belonging to that class.

Precision indicates the reliability of the model when assigning a sample to a particular cataract severity category.

3.7.3 Recall

Recall measures the model's ability to correctly identify samples belonging to a given class. For class c, it represents the proportion of actual class c samples that are correctly recognized by the model. It is defined as:

Recall$_c=\frac{T P_c}{\left(T P_c+F N_c\right)}$     (15)

where, $F N_c$ denotes the false-negative count for class c. Recall is particularly important for the present study because it indicates how effectively the model identifies the less-represented cataract severity categories.

3.7.4 F1-score

The F1-score was calculated separately for each of the four classes to assess class-wise classification performance. Because the Severe category is substantially underrepresented, overall accuracy alone may not adequately reflect minority-class performance. Therefore, class-wise results were interpreted alongside macro- and weighted-averaged measures.

3.7.5 Macro-F1 and weighted-F1

Because of the substantial class imbalance, both macro-averaged and weighted F1-scores were considered. Macro-F1 assigns equal importance to each of the four classes and is therefore useful for examining minority-class performance. In contrast, weighted-F1 accounts for the number of observations in each class and reflects the influence of the overall class distribution.

Both measures were reported to provide complementary perspectives on model performance in the imbalanced four-class classification setting.

3.7.6 Specificity and receiver operating characteristic-area under the curve

Specificity was calculated on a class-wise basis to measure the ability of the model to correctly identify samples that do not belong to a particular class. Receiver operating characteristic-area under the curve (ROC-AUC) was calculated using a one-vs-rest strategy for the four-class classification problem and summarized using the averaging methods reported in the experimental results.

3.7.7 Cohen's Kappa

Cohen's Kappa quantifies the agreement between the predicted and ground-truth class labels while accounting for agreement that may occur by chance.

3.7.8 Matthews Correlation Coefficient

The multiclass Matthews Correlation Coefficient (MCC) was calculated as an additional balanced measure of classification performance based on the complete pattern of correct and incorrect predictions across all four classes.

3.7.9 Confusion matrix

Prediction patterns were analyzed by using the confusion matrix. The diagonal elements depict correct classification of observations, while the off-diagonal elements depict misclassifications between severity categories. This analysis was particularly apt for studying the confusion between Mild, Moderate, and Severe cases.

3.7.10 Statistical and computational evaluation

The same set of 585 independent observations was used to assess all three configurations. Therefore, pairwise comparisons were made with exact McNemar's test obtained from the same observations of test items. The total, trainable, and non-trainable parameters, trainable parameter ratio, model size, inference time, and throughput were used to evaluate computational characteristics.

3.7.11 Interpretability using Gradient-weighted Class Activation Mapping

Class-specific heat maps were then created using Gradient-weighted Class Activation Mapping (Grad-CAM), and the regions of the image that were relevant to the predicted cataract severity were visualized [35]. For qualitative interpretation, the heatmaps were overlaid on the respective fundus photographs. Because the fundus changes that may occur in cataract patients do not always represent a localized retinal lesion, Grad-CAM was not used as a means of clinical localization, but rather as a qualitative interpretability tool.

The present study evaluated three EfficientNet-B0 training configurations for four-class cataract severity classification from fundus images comprising the Normal, Mild, Moderate, and Severe categories. Among the evaluated configurations, the Final Frozen EfficientNet-B0 provided the most favorable overall balance of predictive performance, class-wise behavior, and computational characteristics. It achieved an accuracy of 88.38%, a macro F1-score of 56.33%, and a macro ROC-AUC of 91.60% on the independent patient-level test set.

The results indicate improved recognition of the underrepresented Moderate category compared with the Frozen Baseline configuration, whereas recognition of the severely underrepresented Severe category remained limited. Grad-CAM analysis provided qualitative insight into the image regions associated with model predictions. Pairwise comparison using McNemar's test showed a statistically significant difference between the Severe-Focused and Final Frozen EfficientNet-B0 configurations, whereas no statistically significant difference was observed between the Frozen Baseline and Final Frozen EfficientNet-B0 configurations. The Final Frozen EfficientNet-B0 showed the most favorable overall balance among the evaluated configurations, while recognition of the severely underrepresented Severe category remained limited.

Overall, the findings demonstrate the potential of EfficientNet-B0 for four-class cataract severity classification under the present experimental protocol. However, the limited number of Severe images in the independent test set and the absence of external validation restrict the generalizability of the findings. Future work should evaluate the approach using larger, more balanced, clinically annotated, and externally validated multicenter datasets, with particular emphasis on improving the classification of the underrepresented Severe category.