Improving Classification of Medical Images Using ESRGAN-Based Upscaling and MobileNetV2

Abstract

Medical image classification is an essential component in the development of computer-aided diagnosis systems. One of the main challenges in this field is the limitation of image resolution, which can negatively impact the performance of machine learning models. Low-quality images often lead to a decrease in classification accuracy, particularly in identifying critical features in sensitive cases such as skin and ocular disorders. Therefore, improving image quality is a strategic step that is highly necessary to optimize the overall performance of classification systems. Low-resolution photos are frequently problematic in the medical field when diagnosing skin and eye conditions since they can induce noise and lower the precision of classification algorithms. To overcome this, this research implements the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) method, which is used to perform upscaling, namely increasing the resolution of a low image to a high-resolution image. The research results show that ESRGAN can improve the quality of eye and skin images, as proven by accuracy consistency tests on the two datasets. For image classification, the MobileNetV2 model is used because this model is suitable for eye and skin datasets. Evaluation of the image retrieval system using a high-resolution dataset generated from ESRGAN upscaling demonstrates consistent performance, with a slight accuracy improvement of 1%, reflecting the system's robustness in maintaining reliable results across datasets. In this research, the improvement in visual image quality is also proven by the high Peak Signal-to-Noise Ratio (PSNR) value, so ESRGAN is proven effective in increasing image resolution and clarity for eye medical image datasets and skin images.