Comparative evaluation of classical and machine learning methods for medical image enhancement

Abstract

Medical imaging is critical for diagnostic accuracy, yet raw images often suffer from noise and low contrast. This study provides a comparative evaluation of classical methods, namely the Laplace transform (LT), Sobel operator (SO), and histogram equalization (HE), against a data-driven convolutional neural network (CNN) using the musculoskeletal radiographs (MURA) and Human Metapneumovirus (HMPV) lung computed tomography (CT) datasets. While quantitative analysis shows that HE and SO significantly outperform other methods in isolated contrast enhancement and edge definition, they often introduce artifacts. In contrast, the CNN based approach demonstrates superior detail preservation and entropy, offering a more balanced and adaptive solution for diverse diagnostic requirements. Our findings statistically validate that although classical operators remain highly effective for specific boundary detection tasks, machine learning (ML) frameworks provide the most robust performance for cross-modality image enhancement, bridging the gap between raw data acquisition and clinical interpretation.