Sviluppo e generalizzazione di un modello U-Net per il beamforming e la segmentazione simultanea di immagini ad ultrasuoni = Development and generalization of a U-Net model for simultaneous beamforming and segmentation of ultrasound images

Ultrasound images are widely used as a diagnosis support in the clinical field. This imaging modality has been recognized, over the years, as one of the most adaptable, thanks to its cost-effectiveness and portability, especially if compared to CT and MRI. Nevertheless, US also presents difficult challenges, such as the presence of noise and artifacts that reduce the quality of images, making the diagnosis substantially dependent on the radiologist’s skills and experience. Deep learning approaches have skyrocketed over the last years and have been applied to overcome these limits in different tasks involving ultrasounds, such as for classification, segmentation, and image quality assessment. Recently, the possibility of using deep learning to reconstruct a B-mode image and obtain simultaneously a segmentation mask, utilizing only raw IQ signal, has been demonstrated. This approach can replace the traditional delay-and-sum beamforming algorithm and provide an image with superior quality, by removing the acoustic clutter inside anechoic objects. The DNN architecture was based on a U-Net model with one encoder and two decoders, trained with a dataset containing B-mode images with a single anechoic cyst, created using the Field II simulation program. The aim of this thesis work is to extend the current available knowledge, assessing the effects on the predicted images by training the network using cysts with different degrees of echogenicity (anechoic, iperechoic and ipoechoic), different shapes (circular or bunches of ellipses), different multiplicity and evaluating the outcomes due to the presence of tissue attenuation. The dataset was composed of simulated and experimental ultrasound images. A total of 8192 (6560 training set, 816 validation set and 816 test set) simulated images were generated using Field II on MATLAB® 2019b software. The performances were compared using three different encoder architectures (VGG-13, VGG-16 and VGG-19), and appraising the benefits using a network trained with images which had an automatically determined dynamic range, in contrast to the standard one (60 dB). Particular attention was paid to how the network reconstructed the image from real experimental data, extracted from 24 phantom images and 21 in-vivo carotid images, acquired with a Verasonics® Vantage™ 128 system. For the in-vivo carotid images, the VGG-19 architecture trained with an automatic dR provided a significant performance increase compared to VGG-13 (DSC 0.72±0.22 vs 0.01±0.01, Contrast -5.47±2.84 dB vs -3.68±9.00 dB, SNR 6.29±1.12 vs 8.28±1.08, gCNR 0.34±0.10 vs 0.41±0.13, PSNR 19.01±0.67 dB vs 15.40±0.77 dB). The obtained results show the feasibility of the training of a DNN for the beamforming and segmentation of ultrasound images containing structures with different echogenicity and shapes.