Enhancing Medical Image Classification: A Hybrid Deep Learning Approach with ResNet, Vision Transformer and Capsule Networks

Artificial Intelligence (AI) has proven to hold great promise for early-stage disease diagnosis, providing quick and efficient healthcare solutions. Conventional diagnostic techniques, though useful, tend to lag behind new variations of diseases, and a lot of time and effort are consumed in the process. AI, especially deep learning, provides effective screening and accurate disease prediction from patient reports. The performance of AI-based image analytics in recognizing similar infections has been well documented. Convolutional Neural Networks (CNNs) have come forward as an efficient tool for machine-based disease diagnosis, particularly in medical imaging, offering a robust option for speedy and accurate initial screening. In this research, we discuss the utilization of deep learning frameworks for detection of eye diseases. We compare a baseline CNN’s performance to powerful architectures. Specifically, we probe the role of transfer learning through the utilization of a pre-trained ResNet50 model. Secondly, we inspect the value addition of adding Convolutional Block Attention Module (CBAM) to ResNet50 in augmented feature extraction. In addition, we further our analysis by integrating ResNet50 with Vision Transformers (ViTs) for global context understanding and Capsule Networks for accurate spatial representation, both enhanced with CBAM. Our research evolves from the traditional CNN to an advanced hybrid model, facilitating a thorough evaluation of the growing effectiveness of deep learning in improving accuracy and efficiency in the diagnosis of eye disease, similar to improvements seen in the detection of viral infection.