Abstract

Medical ultrasound imaging is a widely used diagnostic tool in clinical practice, offering several advantages, including high temporal resolution, non-invasiveness, cost-effectiveness, and portability. Despite these benefits, ultrasound modality often suffers from lower image quality compared to other modalities, such as magnetic resonance imaging, which complicates image interpretation and poses diagnostic challenges, even for experienced clinicians. Given its unique advantages, simplifying the interpretation of ultrasound images can profoundly impact the accessibility and affordability of healthcare. This thesis aims to enhance the interpretability of ultrasound images using deep learning (DL)-based approaches on two parallel fronts.
The first front focuses on improving image quality by addressing the phase aberration effect, a primary contributor to the degradation of medical ultrasound images. Phase aberration arises from spatial variations in sound speed within heterogeneous media, introducing artifacts such as blurring and geometric distortions. This effect hinders the accurate representation of tissue structures and complicates clinical interpretation. To tackle this, we propose two novel methods. The first involves training a convolutional neural network (CNN) to estimate the aberration profile from the B-mode image and employing it to compensate for the aberration effects. The second introduces an aberration-to-aberration approach combined with an innovative loss function to train a CNN that directly predicts corrected radio frequency data without requiring ground truth.
The second front focuses on the automatic segmentation of ultrasound images and explores the challenges associated with employing DL-based approaches. Manual segmentation, typically performed by expert clinicians, is time-consuming and prone to human error, and automating this process can simplify the interpretation of ultrasound images. While DL methods have demonstrated considerable potential, ultrasound image segmentation poses unique challenges due to artifacts such as shadowing, reverberation, refraction, phase aberration, and speckle noise. The scarcity of medical data further complicates these challenges, limiting the generalizability and robustness of models in clinical settings. To address these limitations, we investigate the shift-variance problem in CNNs and propose pyramidal blur-pooling layers to mitigate this issue. Furthermore, we tackle domain shift and data scarcity by employing a domain adaptation method and introducing an ultra-fast ultrasound image simulation technique based on frequency domain analysis.