TITLE:
A Practical UNet Denoising Algorithm for Enhanced Malaria Detection in Thick Blood Smear Images
AUTHORS:
Benard Bright
KEYWORDS:
Malaria Diagnosis, Image Denoising, UNet Architecture, Deep Learning, Medical Image Processing, Gaussian Noise, Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Microscopy Images, Noise Reduction, Biomedical Imaging, Image Preprocessing, Data Augmentation, High-Pass Filtering, Fourier Transform, Signal-to-Noise Ratio (SNR), Pixel Intensity Distribution, Image Quality Metrics, Neural Networks, Clinical Image Enhancement
JOURNAL NAME:
Open Access Library Journal,
Vol.12 No.12,
December
24,
2025
ABSTRACT: This paper discusses the task of enhancing malaria detection in thick blood smear images by proposing a UNet-based denoising algorithm. Noise and artifacts in these images can compromise the accuracy of malaria diagnosis. The algorithm, based on the UNet architecture, is developed to remove noise and artifacts, facilitating easier and more accurate identification of malaria parasites. Various preprocessing techniques, including median filters, mean filters, and morphological filters, are explored to mitigate prevalent noise types like speckle, Gaussian, and salt-and-pepper noise. The significance of denoising lies in its potential to minimize misdiagnoses that contribute to false positives and negatives in malaria-related cases, thereby reducing unnecessary drug administration and potential health complications. The proposed UNet denoising algorithm is trained on datasets containing both noisy and clean thick blood smear images. Evaluation against existing denoising methods demonstrates superior performance in terms of denoising quality and malaria detection accuracy. The outcomes reveal the algorithm’s effectiveness in improving the accuracy of malaria diagnosis by effectively removing noise and artifacts from thick blood smear images. The UNet denoising algorithm showed a Structured Similarity Index of 0.92 on average with a minimum SSIM of 0.78 and a maximum SSIM of 0.98. When the images from the dataset with these results were fed into a malaria parasite detection model, model yielded a precision was 0.75, indicating that 75% of the identified “Parasites” are correct, recall of 1.00, meaning that all instances of “Parasites” were correctly identified and an F1-Score of 0.86 demonstrating a balance between precision and recall for the “Parasites” class. This paper underscores the practicality and efficacy of the UNet-based denoising algorithm as a promising solution for enhancing malaria detection in thick blood smear images, offering a significant stride towards more accurate and reliable diagnostics in the fight against malaria.