A.A. Egorchev1, D.E. Chikrin2

1,2 Kazan (Volga Region) Federal University (Kazan, Russia)

1 anton@egorchev.ru, 2 dmltry.kfu@ya.ru

Problem statement. Ultrasound imaging is widely used in medical diagnostics due to its non-invasive nature, real-time capability, and cost-effectiveness. However, the inherent speckle noise, which originates from coherent interference of scattered ultrasound waves, significantly degrades image quality by reducing contrast resolution and obscuring fine anatomical structures. This degradation introduces uncertainty into subsequent quantitative analysis and parameter extraction within information-measurement systems. Classical denoising approaches, including median and Gaussian filters, achieve partial noise suppression but inevitably compromise diagnostically significant details such as tissue boundaries and small structural elements. In recent years, deep learning methods, including convolutional neural networks applied to medical signal processing, have demonstrated superior performance in image restoration tasks by learning complex nonlinear mappings between corrupted and clean image representations.

Aim. To develop a speckle noise suppression method for ultrasound images based on a convolutional autoencoder that performs nonlinear image restoration while preserving diagnostically significant details through a multi-layer convolutional architecture, and to design a threshold-based classification method for noise level estimation using a set of objective image-quality metrics.

Results. A convolutional autoencoder architecture adapted for ultrasound image processing was developed. The encoder compresses the input grayscale image of 1024×1024 pixels through four successive stages of convolutional layers with 128 filters each and max-pooling operations, extracting hierarchical feature representations and reducing spatial dimensions to 64×64 pixels. The decoder reconstructs the denoised image through symmetric upsampling and convolutional layers with a sigmoid activation at the output, minimizing the mean squared error between the original and restored images using the Adam optimizer. The model was trained and evaluated on a publicly available breast ultrasound dataset comprising 780 images across three diagnostic categories: normal, benign and malignant. Restoration quality was assessed using a comprehensive set of full-reference objective metrics: PSNR, MSE, SSIM, ERGAS and UQI. A threshold-based noise level classification method was additionally developed, enabling automatic categorization of images into weak, moderate and strong noise levels based on multi-metric voting analysis. Experimental results on a sample of 1000 images demonstrate that the proposed autoencoder effectively suppresses speckle noise while preserving diagnostically relevant information, achieving PSNR of 34.6 dB and SSIM of 0.91 on representative samples, with an average processing time of 63 ms per image.

Practical significance. The proposed approach provides a robust framework for quality control of the measurement channel and adaptive tuning of processing algorithms in information-measurement systems of medical application. It can also be used for preprocessing of ultrasound data in automated image-analysis pipelines.

Egorchev A.A., Chikrin D.E. Methods for evaluation and suppression of speckle noise in ultrasound measurement data using neural network technologies // Achievements of Modern Radioelectronics. 2026. V. 80. № 9. P. 16−25. DOI: https://doi.org/10.18127/j20700784-202609-02

1. Ortiz S.H.C., Chiu T., Fox M.D. Ultrasound image enhancement: A review. Biomedical Signal Processing and Control. 2012. V. 7. № 5. P. 419–428.
2. Michailovich O.V., Tannenbaum A. Despeckling of medical ultrasound images. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control. 2006. V. 53. № 1. P. 64–78.
3. Gour I., Rajput K. Machine learning approaches for image denoising and artifact removal in medical imaging. International Journal of Intelligent Systems and Applications in Engineering. 2023. V. 11. № 7s. P. 65–71.
4. Mishra D., Chaudhury S., Sarkar M., Soin A.S. Ultrasound image enhancement using structure oriented adversarial network. IEEE Signal Processing Letters. 2018. V. 25. № 9. P. 1349–1353.
5. Karaoğlu O.R., Bilge H.Ş., Uluer I. Removal of speckle noises from ultrasound images using five different deep learning networks. Engineering Science and Technology, an International Journal. 2022. V. 29. P. 101030.
6. Gondara L.Y. Medical image denoising using convolutional denoising autoencoders. Proceedings of IEEE International Conference on Data Mining Workshops (ICDMW). 2016. P. 241–246.
7. Lee W.H., Ozger M., Challita U., Sung K.W. Noise learning-based denoising autoencoder. IEEE Communications Letters. 2021. V. 25. № 9. P. 2983–2987.
8. Oliveira-Saraiva D., Mendes J., Leote J., Gonzalez F.A., Garcia N., Ferreira H.A., Matela N. Make it less complex: Autoencoder for speckle noise removal – application to breast and lung ultrasound. Journal of Imaging. 2023. V. 9. № 10. Art. 217.
9. Al-Dhabyani W., Gomaa M., Khaled H., Fahmy A. Dataset of breast ultrasound images. Data in Brief. 2020. V. 28. Art. 104863.
10. Kokil P., Sudharson S. Despeckling of clinical ultrasound images using deep residual learning. Computer Methods and Programs in Biomedicine. 2020. V. 194. P. 105477.
11. Wang Z. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing. 2004. V. 13. № 4. P. 604–606.
12. Javan F.D., Samadzadegan F., Mehravar S., Toosi A., Khatami R., Stein A. A review of image fusion techniques for pan-sharpening of high-resolution satellite imagery. ISPRS Journal of Photogrammetry and Remote Sensing. 2021. V. 171. P. 101–117.
13. Irmak E., Ertas A.H. A review of robust image enhancement algorithms and their applications. Proceedings of IEEE Smart Energy Grid Engineering (SEGE). 2016. P. 371–375.
14. Kompanets I., Zalyapin N. Methods and devices of speckle-noise suppression. Optical and Photonic Journal. 2020. V. 10. № 10. P. 219–250.
15. Bianco V., Memmolo P., Leo M., Montresor S., Distante C., Paturzo M., Ferraro P. Strategies for reducing speckle noise in digital holography. Light: Science and Applications. 2018. V. 7. № 1. Art. 48.
16. Ponomarenko N.N., Lukin V.V., Zelenskij A.A., Egiazaryan K.O., Astola Ya., Karli M., Battisti F. TID2008 – A database for evaluation of full-reference visual quality assessment metrics. Uspexi sovremennoj radioe`lektroniki. 2009. № 10. S. 30–45. (in Russian).
17. Nepomnyashhij O.V., Moxammed M.I. Al.S., Xajdukova V.N., Sharova A.V., Aleksandrovich K.S. Convolutional neural network and deep machine learning for ECG analysis. Uspexi sovremennoj radioe`lektroniki. 2024. T. 78. № 3. S. 70–77. DOI: 10.18127/j20700784-202403-08. (in Russian).