Multi-Domain Medical Image Classification Using EfficientNet-B3

Abstract

Medical image analysis has seen considerable progress with the integration of deep learning approaches, especially convolutional neural networks (CNNs), which are known to be successful in extracting relevant spatial information from complex biomedical data. However, most existing models are generally developed for a particular application and are not flexible enough to be used on multiple disparate datasets in a single system. In real-world medical applications, ECE images for cardiac analysis, EEG-based graphical representations for seizure analysis, and facial images for emotion analysis are quite disparate in terms of structural information and feature distributions, which make it quite challenging for existing CNN models to generalize across domains without resorting to multiple models or repeating training. These application-specific models are quite computationally intensive and less efficient in terms of deployment. To overcome these hurdles, this study proposes a unified classification platform based on the EfficienyNet-B3 model, which is capable of handling multiple biomedical imaging applications using a single backbone model. The proposed platform incorporates a dynamic key-driven selection module that enables runtime switching between applications without requiring the loading of multiple independent models, thus overcoming redundancy while preserving excellent predictive accuracy and scalability for real-world applications.