Hybrid Wavelet-Deep Learning Framework for Fluorescence Microscopy Images Enhancement

Abstract

Fluorescence microscopy is a powerful tool for visualizing cellular structures, but it faces challenges such as noise, low contrast, and autofluorescence that can hinder accurate image analysis. To address these limitations, we propose a novel hybrid image enhancement method that combines wavelet-based denoising, linear contrast enhancement, and convolutional neural network-based autofluorescence correction. Our automated method employs Haar wavelet transform for noise reduction and a series of adaptive linear transformations for pixel value adjustment, effectively enhancing image quality while preserving crucial details. Furthermore, we introduce a semantic segmentation approach using CNNs to identify and correct autofluorescence in cellular aggregates, enabling targeted mitigation of unwanted background signals. We validate our method using quantitative metrics, such as signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR), demonstrating superior performance compared to both mathematical and deep learning-based techniques. Our method achieves an average SNR improvement of 8.5 dB and a PSNR increase of 4.2 dB compared with the original images, outperforming state-of-the-art methods such as BM3D and CLAHE. Extensive testing on diverse datasets, including publicly available human-derived cardiosphere and fluorescence microscopy images of bovine endothelial cells stained for mitochondria and actin filaments, showcases the flexibility and robustness of our approach across various acquisition conditions and artifacts. The proposed method significantly improves fluorescence microscopy image quality, facilitating more accurate and reliable analysis of cellular structures and processes, with potential applications in biomedical research and clinical diagnostics.