A Novel Transformer Model With Multiple Instance Learning for Diabetic Retinopathy Classification

Aim:

      To design and implement an advanced diagnostic system for retinal disease classification, combining state-of-the-art feature extraction and classification models for superior accuracy.

Abstract:

      The accurate and efficient diagnosis of retinal diseases such as diabetic retinopathy, glaucoma, and hypertensive retinopathy is essential to prevent vision impairment and blindness. Current diagnostic methods often face limitations due to low resolution, computational overhead, and reduced sensitivity to subtle abnormalities. To address these issues, we propose a system that employs RetFound, a pre-trained model designed for high-resolution retinal image feature extraction, combined with ResNet50v2, known for its robust classification capabilities.

        By training the extracted features on ResNet50v2, the system improves accuracy and sensitivity across all four classes, including healthy retinopathy. Deployment is realized through a Streamlit application, enabling real-time predictions and ease of use. Future work will incorporate additional retinal diseases and advanced deep learning models such as Generative Adversarial Networks (GANs) to enhance diagnostic capabilities further.

Introduction:

     Retinal diseases are a major cause of blindness globally, with diabetic retinopathy, glaucoma, and hypertensive retinopathy contributing significantly to this burden. Early detection is crucial, as many retinal abnormalities manifest subtly before causing irreversible damage. Current diagnostic approaches are time-consuming, resource-intensive, and often limited by the expertise required. With advancements in artificial intelligence and deep learning, automated diagnostic systems now offer a promising solution.

    However, existing systems face challenges like low resolution, computational inefficiency, and inadequate feature extraction. This study aims to overcome these limitations by integrating the pre-trained RetFound model for feature extraction and the ResNet50v2 architecture for classification. The proposed system also ensures real-time usability via Streamlit, promoting accessibility in clinical and remote settings.

Problem Definition:

      Accurate classification of retinal diseases is challenging due to the complexity of retinal structures and the overlapping features among diseases. For example, lesions in diabetic retinopathy and glaucoma often share visual similarities that can confuse classification models. The high resolution of retinal images adds computational challenges, particularly for deep learning models like transformers and CNNs, which are often constrained by input size limitations. Furthermore, the lack of user-friendly deployment mechanisms prevents widespread adoption of AI-based systems in practical scenarios. Addressing these issues requires a system that balances accuracy, computational efficiency, and usability without sacrificing the diagnostic performance required for medical applications.

Existing System:

       Traditional retinal disease classification systems rely heavily on convolutional neural networks (CNNs) like ResNet and InceptionV3. While these models achieve reasonable performance, they often struggle with high-resolution retinal images due to their limited capacity to process global features. Transformer-based models such as ViT (Vision Transformer) have shown promise in improving classification accuracy by leveraging self-attention mechanisms. However, they require image resizing to 224×224, which leads to a significant loss of detail, especially for small but critical features like microaneurysms or hemorrhages. systems like TMIL (Transformer-based Multiple Instance Learning) attempt to address this by segmenting images into smaller patches for feature extraction. However, such systems remain computationally demanding, limiting their real-time applicability in clinical settings. Moreover, these approaches often require extensive pre-training on large datasets, which may not always be available for medical imaging.

Disadvantages:

1. Loss of Resolution: Existing systems like ViT reduce image sizes, which results in critical feature loss, particularly for small lesions. This affects sensitivity and specificity.
2. Computational Demands: High-resolution images significantly increase computational requirements, making real-time processing difficult in clinical environments.
3. Dataset Limitations: Most systems rely on large-scale pre-training to improve accuracy, which is impractical for rare diseases where labeled datasets are sparse.
4. Limited Deployment Usability: Many systems lack intuitive interfaces for end-users, reducing their accessibility in practical scenarios.
5. Reduced Sensitivity: Despite improvements, systems often show false negatives, particularly for subtle disease stages. This reduces the reliability of diagnostic outputs.
6. Overfitting Risks: With limited datasets, models tend to overfit, leading to reduced generalization to unseen cases.

Proposed System:

       The proposed system integrates RetFound for feature extraction with ResNet50v2 for classification. RetFound is a pre-trained model optimized for high-resolution medical imaging, capable of identifying subtle features such as microaneurysms and exudates. Extracted features are input into ResNet50v2, which has been fine-tuned for classifying retinal images into diabetic retinopathy, glaucoma, hypertensive retinopathy, and healthy retinopathy. The combination leverages RetFound’s extraction precision and ResNet50v2’s robust classification performance. Deployment in a Streamlit application ensures a user-friendly, real-time diagnostic experience. The system is trained using publicly available retinal image datasets, employing augmentation techniques to overcome data scarcity. By maintaining high resolution throughout the pipeline, the system preserves critical diagnostic details while achieving superior accuracy and reduced computational demands.