Enhancing Breast Ultrasound Diagnostics Using GANs and Guided Filtering
Taher Slimi, Anouar Ben Khalifa · 2025
In ultrasound imaging, speckle noise presents a significant challenge, undermining image clarity and the accuracy of diagnostic interpretations. To address this critical issue, We develop a new method for improving the quality of ultrasound images using a Denoising Generative Adversarial Network complemented by a Guided Filter (DNGAN-GF). This method uses a neural network architecture to learn how to generate denoised images from noisy inputs while preserving important anatomical contours and details. Our innovative approach effectively transforms noisy images into remarkably clear visual representations. The DNGAN learns to capture the characteristics of clean image data, while the GF refines the results, ensuring the fidelity of anatomical structures. Qualitative assessments conducted by experts highlight the exceptional sharpness of contours and anatomical features, making anomaly interpretation more intuitive and accessible. This improvement not only assists radiologists in their evaluations but also has the potential to reduce diagnostic errors. Quantitative analyses reveal substantial increases in PSNR and SSIM metrics, confirming the superiority of our proposed DNGAN-GF method compared to conventional techniques. By significantly enhancing the visibility of anatomical details, our proposed solution can improve diagnostic accuracy and redefine standards in medical imaging, ultimately elevating the quality of patient care.