• cnn
• denoising
• keras
• speckle noise
• tensorFlow
• u-net
• ultrasound

In medical diagnosis, ultrasound has become a well-established technique due to its numerous advantages, including its non-invasive nature, lack of ionizing radiation, affordability, ability to generate real-time images, and possibility to improve the image quality via proper processing techniques.
Ultrasound imaging faces various challenges, such as equipment acquisition noise, ambient noise, presence of surrounding tissues and organs, and anatomical factors. However, considering the numerous advantages it offers, the importance of devising image denoising techniques becomes evident. Denoising is thus a significant problem in image processing, aiming to restore a clean image from the original noisy one.
The so-called "speckle" noise is the primary source of degradation in ultrasound images; it is multiplicative by nature, and is influenced by the underlying signal, making it an inherent characteristic of ultrasound imaging. Speckle noise tends to degrade sharpness and minute details, reducing resolution and contrast: consequently, this noise may adversely affect the diagnostic value of ultrasound images. Therefore, the removal of speckle noise is considered a crucial preprocessing step. Over the years, various methods have been proposed for this purpose; in this work, the method under investigation adopts a deep learning approach.
This thesis project aims to develop a tool that can be continuously updated to find the best denoising method: with this purpose, the project aims to translate the previous version of a neural network tool written with the TensorFlow 1.0 library and found on GitHub into a new implementation using the TensorFlow 2.0 library, with the high-level API Keras, allowing for greater control over the code.
The final goal of this thesis is the development of a flexible and comprehensible tool for the exploration of different network architectures and approaches. The tool will offer options for sample modification, such as adding simulated noise, applying data augmentation, incorporating methods against overfitting, and more, and users can also make custom modifications to the tool with a high degree of flexibility.
To compare the effectiveness of this tool with a previous version, several experiments have been carried out and reported. Specifically, all the results have been obtained using the Noise2Noise approach and U-Net as the neural network architecture. The U-Net architecture shows a strong capability to reconstruct the major features of the image and to recognize complex patterns, with remarkable denoising power. On the other hand, the Noise2Noise technique relies on a convenient training process by using datasets consisting of pairs of only noisy images, thus with no need for a separate dataset of clean images, which is challenging to obtain in the biomedical field due to the rarity of certain pathologies and concerns surrounding patient privacy, which impose restrictions on access to medical images.
The datasets used for training and subsequent testing have been provided by Esaote, a leading company in medical imaging. The objective was to compare the results obtained during the testing phase using both the previous implementation and the new implementation of the tool. By evaluating and comparing the outcomes, we aimed to assess the performance and effectiveness of the updated implementation in denoising ultrasound images.