Explainable Diabetic Retinopathy Classification Using an Autonomous Learning Multi-Model (ALMMo-0) Classifier with Transfer Learning

Diabetic retinopathy (DR) is a leading cause of preventable blindness among diabetic patients, necessitating early detection and treatment. Manual diagnosis requires significant time and resources, prompting the development of automated detection and classification methods using deep learning techniques [7]. These approaches analyze retinal fundus images to detect blood vessels, hemorrhages, and other DR-related features. Various machine learning algorithms and deep learning models, have been employed to classify DR stages with high accuracy.

The autonomous learning multi-model classifier of 0-order (ALMMo-0) is a noniterative, data-driven classifier that automatically extracts data clouds and forms, for each class, sub-classifiers based on fuzzy rules [8]. While originally parameter-free, a new approach introduces an initial radius hyper-parameter, allowing users to choose between accuracy and complexity [9]. The ALMMo-0 system has been extended to first-order (ALMMo-1) and adapted for multi-class classification tasks, demonstrating flexibility and comparable performance to benchmark methods [10]. Both ALMMo-0 and ALMMo-1 systems have shown high accuracy and efficiency in classification and regression tasks, with the ability to learn from streaming data and self-evolve their structure. These characteristics make ALMMo systems attractive solutions for various real-world applications, offering a balance between performance and adaptability.

The xDNN model achieves a high accuracy of 99.7% on the APTOS-2019 dataset, emphasizing the importance of interpretability in clinical applications [11].

One study reported a deep learning model achieving 94% sensitivity and 98% specificity in DR detection [10]. These automated systems show promise in reducing vision loss by enabling timely referrals to ophthalmologists for further evaluation and treatment [12].

Combining CNNs with techniques like Adaptive Gabor Filters and Random Forests has improved classification accuracy to nearly 98% [13]. Recent models utilize attention mechanisms and vision transformers to enhance feature extraction, achieving accuracies of 99.63% [14].

Transfer learning has been widely used for DR detection. The work presented in the study [15] proposes a model for DR detection based on transfer learning. Bodapati et al. [16] combine feature extraction and transfer learning techniques. Bhardwaj et al. [17] developed a deep learning model to distinguish DR disease identification and its grading using a transfer learning approach. Pour et al. [18] performed feature extraction and classification in DR detection by using EfficientNet.

Jena et al. [19] proposed a novel approach for DR screening using asymmetric deep learning features, achieving 98.6% accuracy on the APTOS dataset and 91.9% on the MESSIDOR dataset. Nur-A-Alam et al. [20] introduced an automated technique for classifying retinal fundus images into DR and normal states using feature fusion, achieving a detection accuracy of 95.75%. Incir and Bozkurt [21] used K-Means clustering for lesion segmentation and pretrained models like EfficientNetV2-M, achieving 95.16% accuracy.

Omer [22] presented a computer-aided screening system (DREAM) utilizing a bilayered neural network for classifying DR severity, achieving 98.5% accuracy on 6,332 fundus images. Akhtar et al. [23] proposed a binary classification framework for DR detection using Transfer Learning, achieving a test accuracy of 97.82% with an image dataset from APTOS-2019. In the reference [24], machine learning algorithms such as logistic regression, naive bayes (NB), support vector machine (SVM) and random forest are used for DR detection and classification.

Table 1. Summary of models and results obtained by related works

Costaner et al. [25] developed a machine learning-based method for DR detection using local binary pattern (LBP) and wavelet transform, achieving 95.59% accuracy, 96% precision, and 97.96% recall with SVM classification. Table 1 summarizes the related works for the DR automatic detection tasks.

Recent research on Transformer-based architectures for diabetic retinopathy (DR) classification has demonstrated impressive results, particularly in improving feature representation and global contextual understanding. For instance, Li and Huang [26] proposed a vision transformer (ViT)-based model that achieved an accuracy of 93.8% and an AUC of 0.97 on the EyePACS dataset, showing strong robustness in detecting different DR severity levels. Similarly, Dosovitskiy [27] highlighted the superior generalization ability of Transformer backbones over CNNs like ResNet50 and VGG16, reporting state-of-the-art results in image recognition tasks with accuracies exceeding 90% in medical imaging benchmarks. In another study, Xu and Wang [28] employed a Swin Transformer-based hierarchical network that reached 95.2% accuracy and an F1-score of 0.94 on the APTOS 2019 dataset, particularly excelling in identifying subtle lesion regions and inter-class boundaries.

Current research on detecting and classifying diabetic retinopathy (DR) by using explainable methods reveals several critical gaps that hinder advancements in accurate diagnosis and treatment. While recent studies have made significant strides using deep learning, AI technologies, and explainable AI (XAI), the integration of these methods into practical clinical applications remains underexplored. There is a pressing need for comprehensive methodologies that effectively integrate various components to address these challenges. Key gaps include:

• Lack of Comprehensive Explainability. Many existing models, such as those utilizing Concept Activation Vectors (CAVs) and Concept Bottleneck Models (CBMs), have not been thoroughly evaluated for their interpretability in clinical contexts [29, 30]. This limits their adoption in real-world healthcare settings.
• Need for Intuitive Explanations. Medical professionals require explanations that align with their clinical understanding, yet current XAI methods often fall short in providing clear, actionable insights into model decisions [31]. This gap reduces the trust and usability of AI systems in clinical practice.
• Underdeveloped Hybrid Frameworks. Hybrid approaches combining techniques like fuzzy logic and explainable neural networks are still in their early stages and require further development to improve accuracy, robustness, and user trust [32].
• Limited Generalizability. Many studies rely on specific datasets, which restricts the generalizability of findings across diverse populations and clinical settings [33]. This makes difficult the application of models in real-world scenarios.
• Challenges with Synthetic Data. While synthetic data generation shows promise for augmenting training datasets, it requires more rigorous validation to ensure robustness and reliability in real-world applications [34].
• Classification Accuracy Issues. Many models struggle with false positives, misclassifying healthy images as diseased, which can lead to unnecessary interventions [35].
• Additionally, there is insufficient emphasis on extracting and utilizing morphological features, such as lesion shape and texture, to improve classification accuracy [36].

To address these gaps, future research should focus on developing comprehensive, intuitive, and clinically relevant frameworks that balance accuracy and explainability. This perspective highlights a potential trade-off between model performance and the clarity of explanations provided to clinicians, underscoring the need for a balanced approach in future research.

The general architecture of our proposal consists of several key components, as shown in Figure 1. These components work together to process empirical data, generate fuzzy rules, and optimize models using the EDA framework.

Figure 1. ALMMo-0 system: general architecture

• Input Data. The raw empirical data collected from the system or environment.
• Preprocessing. The stage where data is cleaned, normalized, and prepared for analysis.
• Feature extraction. The stage where relevant, meaningful and discriminative features are extracted from images.
• AnYa FRB System. The fuzzy rule-based system that generates adaptive rules using data clouds.
• Output Model. The final optimized model ready for deployment or further analysis.

4.1 Preprocessing

Pre-processing fundus images is a crucial step used to reduce noise and inconsistencies from different imaging devices and environments. Techniques like resizing, cropping, contrast adjustment, normalization, and data augmentation applied to enhance image quality. This guarantees classification models emphasis on key features, improving accuracy and robustness in term of classification performance. Particularly in diabetic retinopathy detection, it leads to more reliable and consistent medical image analysis.

Diabetic retinopathy datasets frequently contain fundus images with different resolutions and aspect ratios, sometimes containing black space. To normalize input sizes, cropping image is applied to eliminate useless areas. This ensures images with fixed resolution, which permit to enhancing classification model performance. In addition, CLAHE (Contrast Limited Adaptive Histogram Equalization) method [38] efficiently enhances image quality by improving low-contrast areas. It highlights lesions in fundus images (FIs), making medical image analysis more reliable.

Also, CLAHE enhances local contrast, making subtle details more visible in regions where there are significant variations of intensity levels. One main parameter of CLAHE method is the clip limit: It regulates the contrast adjustment process. Besides, the clip limit parameter plays an important role in balancing image clarity with preservation of details. This user-defined value modifies the histogram to prevent excessive distortion. As a result, good tuning guarantees effective improvement without over-amplifying noise or artifacts.

Data augmentation generates diverse training samples through transformations such as rotation, flipping, scaling, and brightness adjustment, thereby improving model robustness and generalizability.

Finally, normalization standardizes pixel values, scaling them to a consistent range (e.g., [0, 1] or zero mean and unit variance), which stabilizes training and ensures faster convergence.

Together, these preprocessing steps: Circle cropping, CLAHE, data augmentation and normalization, create a robust foundation for accurate and reliable DR detection and classification. Figure 2 shows some examples of original and preprocessed images.

Figure 2. Examples of some preprocessed and original fundus images, associated with their respective classes

4.2 Feature extraction

In computer vision, obtaining relevant features from traits plays a vital role in tasks like object detection, content-based retrieval, and image classification. Deep learning has revolutionized this process by offering advanced methods for feature extraction, particularly leveraging pre-trained CNNs, a widely used approach is transfer learning which enables the adaptation of knowledge from pre-trained models to new applications. Rather than constructing a deep neural network from the ground up, we can utilize a model trained on extensive datasets like ImageNet and fine-tune to enhance performance on specific tasks This approach is especially useful when working with smaller datasets or when computational resources are limited. Transfer learning has been successfully applied in various medical diagnostics, such as developing a cloud-based solution for liver cancer detection using deep learning and classifying cancer from DNA microarray data with genetic algorithms and case-based reasoning. These examples demonstrate how transfer learning can enhance the adaptability and effectiveness of models across different healthcare applications.

One well-known network frequently applied in transfer learning is VGG16, which achieved prominence during the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) due to its impressive accuracy. Developed by the Visual Geometry Group at the University of Oxford, VGG16 is a deep convolutional neural network featuring 16 layers and utilizing small 3´3 convolutional filters consistently. It’s simple yet effective architecture, along with strong performance on ImageNet, has made it a widely adopted choice for numerous computer vision applications. The VGG16 architecture includes several convolutional layers followed by maxpooling layers, leading up to fully connected layers. After the final convolutional layer, which produces a 7´7´512 tensor, the output is flattened into a single-dimensional vector of length 25,088. This vector is then passed through fully connected layers, where it is reduced to a 1´4096-dimensional vector through matrix multiplication and a ReLU activation function.

In transfer learning, VGG16 functions as a feature extractor by retaining its convolutional layers and removing the fully connected layers. This adaptation allows the network to process images and generate a 1´4096-dimensional feature vector. The process involves image pre-processing, passing it through the modified network, and extracting meaningful features. This method enables efficient feature extraction without requiring extensive retraining. Typically, VGG16 uses weights pre-trained on the ImageNet dataset, which captures a broad range of visual features useful for various tasks. If the application domain differs significantly from ImageNet, additional domain-specific pretraining or fine-tuning may be required. However, in many cases, the default ImageNet weights suffice for feature extraction, unless the domain images are vastly different. In our experience, using the ImageNet weights yielded the best results, likely due to the diverse and rich feature representations learned from the extensive ImageNet dataset.

4.3 ALMMo-0 classifier

This section briefly recalls the main notions related to the 0-order AnYa Fuzzy Rule-Based (FRB) system and the EDA estimator. The AnYa FRB system is a type of fuzzy rule-based model that uses data clouds to represent rules, eliminating the need for predefined membership functions. This makes the system highly adaptive and capable of handling non-linear and dynamic data. The EDA estimator, on the other hand, is a computational tool used within the EDA framework to estimate parameters and optimize models based on empirical data. The AnYa Fuzzy rule-based system and the EDA estimator provide Together a flexible and efficient approach for modeling complex systems.

4.3.1 0-Order AnYa fuzzy rule-based system

The ALMMo-0 classification consists of a collection of AnYa fuzzy rules [8]. Unlike the commonly used Mamdani and Assilian [39], Zadeh [40] and Takagi and Sugeno [41] fuzzy rule-based (FRB) systems, in an AnYa fuzzy rule, the antecedent is simplified into a vector representing the focal points corresponding to the different data clouds. The concept of data clouds refers to clusters of data samples with shared characteristics, organized around focal points similar to Voronoi tessellation [42]. In the AnYa approach, the data clouds as well as their focal points serve as the foundation for the antecedent, i.e., the IF part, of the fuzzy rule. A zero-order AnYa fuzzy rule is formulated as follows:

Rule $i$ : IF $x \approx x_i^*$ $THEN \,Label$${ }_i$            (1)

where, $x_i^*$ is the focal point of the $i^{\text {th }}$ cloud; $L a b e l_i$ is the corresponding label. When classification is considered, inference in the 0-order AnYa rule of is done based on the principle "winner takes all".

4.3.2 EDA estimator

In the present paper, the EDA framework, and especially the unimodal density, is used as the main estimator to autonomously reveal global properties from observed data. We define the dataset or data stream in the Euclidean space $\mathbb{R}^d$ as $\left\{x_1, x_2, \ldots, x_k\right\}$, where subscripts denote time instances of data observation. For simplicity, Euclidean distance is used in the mathematical formulation, though other distance metrics can also be applied. The unimodal density of the i^th data sample at the k^th time instance is computed as:

$D_k\left(x_i\right)=\frac{1}{1+\frac{\left\|x_i-\mu_k\right\|^2}{\sigma_k^2}}=\frac{1}{1+\frac{\left\|x_i-\mu_k\right\|^2}{X_k-\left\|\mu_k\right\|^2}}$              (2)

where, $\mu_k$ is the mean of all the samples computed at the $k^{t h}$ time instance and $X_k$ is the average scalar product: $\sigma_k^2=X_k- \left\|\mu_k\right\|^2$. It is worth noting that in the case of Euclidean distance, the unimodal density has the form of a Cauchy function, even if there is no assumption that the distribution is a Cauchy distribution.

For efficient streaming data processing, recursive computation plays a fundamental role in optimizing memory usage and computational performance. The values of $\mu_k$ and $X_k$ are updated by using Eqs. (3) and (4), that recursively compute the unimodal density without explicit loops:

$\mu_k=\frac{k-1}{k} \mu_{k-1}+\frac{1}{k} x_k ; \quad \mu_1=x_1$              (3)

$X_k=\frac{k-1}{k} X_{k-1}+\frac{1}{k}\left\|x_k\right\|^2 ; \quad X_1=\left\|x_1\right\|^2$            (4)

4.3.3 Overview of multiple model architecture

This architecture utilizes multiple sub-classifiers to process incoming data samples within a classification framework. We evaluate every new data sample, $x_k$, by all available subclassifiers. Each sub-classifier $i$ produces a confidence score, $\lambda_i$, representing the probability that $x_k$ belongs to a particular class. The final classification is determined using a "winner-takes-all" approach, where $x_k$ is assigned to the class with the highest confidence score.

Label $=\operatorname{argmax}_{i=1,2, \ldots, R}\left(\lambda_i\right)$              (5)

This multiple-model approach enhances the classifier’s capacity for handling complex problems by combining the strengths of each sub-classifier as shown in Figure 3.

Figure 3. A conceptual framework diagram of a multiple-model classifier

4.3.4 Learning stage in the ALMMo-0 classifier

During the learning phase, we only update the AnYa fuzzy rule-based (FRB) rules relaying on new data sample’s class with normalizing these new samples:

$x_k \leftarrow \frac{x_k}{\left\|x_k\right\|}$              (6)

In the case of high-dimensional data, this normalization improves the classifier's performance. Let $x_k^i$ be a new data sample from class $i$. We update $\mu_{k-1}^i$, which denotes the class's global mean, to a new mean $\mu_k^i$. Since each data sample is normalized, the update of the average scalar product is not necessary.

The found focal points of class $i$, denoted as $x_j^{* i}$ for $j= 1,2, \ldots, F_i$ (where $F_i$ is the number of focal points) as well as the unimodal densities of the new data sample $x_k^i$ are computed using the following Eq. (7):

Dendity $=f\left(x_k^i, x_j^{*_i}\right)$            (7)

This density computation helps the classifier to effectively adapt itself to changing data distributions, particularly in high-dimensional spaces.

To determine if $x_k^i$ should create a new data cloud or a new rule, the following condition (Condition 1) is checked:

$\operatorname{IF}\binom{\left(D_k\left(x_k^i\right)>\max _{j=1,2, \ldots, F_i}\left(D_k\left(x_j^{* i}\right)\right)\right)}{\operatorname{OR}\left(D_k\left(x_k^i\right)>\max _{j=1,2, \ldots, F_i}\left(D_k\left(x_j^{* i}\right)\right)\right)}$ THEN Add $x_k^i$ as a novel focal point            (8)

In the case where Condition 1 holds, a new fuzzy rule or data cloud is constructed and associated with $x_k^i$. The adaptation of the parameters of this new data cloud is done as follows:

$\left\{\begin{array}{l}F^i \leftarrow F^i+1 \\ x_{F^i}^{* i} \leftarrow x_k^i \\ M_{F^i}^{* i} \leftarrow 1 \\ r_{F^i}^{* i} \leftarrow r_0\end{array}\right.$            (9)

where, $M_{F^i}^{* i}$ is the number of members in the data cloud, $r_{F^i}^{* i}$ is the radius of the influence area, and $r_o$ is a small stabilizing value for initializing new data clouds, set by $r_0= \sqrt{2\left(1-\cos \left(15^{\circ}\right)\right)}$.

In the case where Condition 1 does not hold, Eq. (10) is used to identify the nearest data cloud to $x_i^k$:

$x_N^{* i}=\operatorname{argmin}_{j=1, \ldots, F_i} x k i-x j * i$               (10)

If Condition 2 is verified $\left(\left\|x_k^i-x_N^{* i}\right\| \leq r_N^{* i}\right)$, then $x_i^k$ is assigned to the nearest data cloud. Besides, the following Eq. (11) shows how the meta-parameters of the nearest data cloud are updated:

$\left\{\begin{array}{c}x_N^{* i} \leftarrow \frac{M_N^{* i}}{M_N^{* i}+1} x_N^{* i}+\frac{1}{M_N^{* i}+1} x_k^i \\ M_N^{* i} \leftarrow M_N^{* i}+1 \\ r_N^{* i} \leftarrow \sqrt{0.5\left(r_N^{* i}\right)^2+\left(1-\left\|x_N^{* i}\right\|^2\right)}\end{array}\right.$                (11)

In the case where Condition 2 does not hold, $x_k^i$ gives rise to a new data cloud using the parameters defined in Eq. (8). Notice that, for the next cycle, no change is performed on the parameters of data clouds without new members.

Algorithm 1 summarizes the previous steps of the learning stage.

4.3.5 Validation stage

 During validation, each sample is given as input to the different AnYa FRB sub-classifiers that correspond to our $C$ classes. Each AnYa FRB rule $j($ for $j=1,2, \ldots, R)$ generates a confidence score as follows:

$\lambda_j=e^{-\frac{1}{2}\left\|x_k-x_j^*\right\|^2}$             (12)

After all R rules have generated their scores, the rule with the highest confidence score is selected based on the “winner takes all” principle. This assigns the appropriate label to the validation data sample.