Recently, research has been focusing on the development of artificial intelligence ultrasound beamforming methods to improve the contrast and resolution of B-mode images. In this work, we propose an innovative beamforming domain transfer method using a generative adversarial network (GAN). The GAN takes as input a plane-wave (PW) delay and sum (DAS) image and generates an image as if it had been acquired using the focused modality and reconstructed with the filtered Delay Multiply and Sum (F-DMAS) beamforming technique. A Verasonics Vantage 256 system (L11-5v linear array) was used to acquire 560 (480 and 80 for train and test set, respectively) in-vivo musculoskeletal US images. Images were acquired on five muscles (gastrocnemius lateralis, gastrocnemius medialis, vastus lateralis, vastus medialis, and biceps) on both sides of 14 healthy volunteers (50% female). RF data were acquired both in plane-wave (PW) and focused mode and beamformed using the UltraSound ToolBox (USTB). The DAS beamforming method was employed for PW data, whereas the focused data were reconstructed using F-DMAS. Various dynamic ranges (dR) were employed to create the final 8-bit PW DAS images (dR = 55, 65, 75, 85 dB) while an automatic dR was employed to optimize focused F-DMAS images. A Pix2Pix GAN architecture was designed to formulate the task of beamforming as the translation from one domain (PW DAS image) to another (focused F-DMAS image). Our GAN employed a UNet as the generator and a 3-layer fully convolutional PatchGAN as the discriminator. The proposed GAN architecture shows promising results, generating a GAN image comparable to the F-DMAS image, i.e., in terms of SSIM (0.5183 +/- 0.0437 and 0.5152 +/- 0.0519 for GAN images vs DAS images and F-DMAS images vs DAS images). Overall, our GAN enhances image quality and simulates focused F-DMAS beamforming starting from a PW DAS image without needing to access the raw RF data, which is typically unavailable with clinical ultrasound devices.

Ultrasound Image Beamforming Optimization Using a Generative Adversarial Network / Seoni, S; Salvi, M; Matrone, G; Meiburger, Km. - ELETTRONICO. - (2022), pp. 1-4. (Intervento presentato al convegno IEEE International Ultrasonics Symposium (IUS) tenutosi a Venice, Italy nel 10-13 October 2022) [10.1109/IUS54386.2022.9957306].

Ultrasound Image Beamforming Optimization Using a Generative Adversarial Network

Seoni, S;Salvi, M;Meiburger, KM
2022

Abstract

Recently, research has been focusing on the development of artificial intelligence ultrasound beamforming methods to improve the contrast and resolution of B-mode images. In this work, we propose an innovative beamforming domain transfer method using a generative adversarial network (GAN). The GAN takes as input a plane-wave (PW) delay and sum (DAS) image and generates an image as if it had been acquired using the focused modality and reconstructed with the filtered Delay Multiply and Sum (F-DMAS) beamforming technique. A Verasonics Vantage 256 system (L11-5v linear array) was used to acquire 560 (480 and 80 for train and test set, respectively) in-vivo musculoskeletal US images. Images were acquired on five muscles (gastrocnemius lateralis, gastrocnemius medialis, vastus lateralis, vastus medialis, and biceps) on both sides of 14 healthy volunteers (50% female). RF data were acquired both in plane-wave (PW) and focused mode and beamformed using the UltraSound ToolBox (USTB). The DAS beamforming method was employed for PW data, whereas the focused data were reconstructed using F-DMAS. Various dynamic ranges (dR) were employed to create the final 8-bit PW DAS images (dR = 55, 65, 75, 85 dB) while an automatic dR was employed to optimize focused F-DMAS images. A Pix2Pix GAN architecture was designed to formulate the task of beamforming as the translation from one domain (PW DAS image) to another (focused F-DMAS image). Our GAN employed a UNet as the generator and a 3-layer fully convolutional PatchGAN as the discriminator. The proposed GAN architecture shows promising results, generating a GAN image comparable to the F-DMAS image, i.e., in terms of SSIM (0.5183 +/- 0.0437 and 0.5152 +/- 0.0519 for GAN images vs DAS images and F-DMAS images vs DAS images). Overall, our GAN enhances image quality and simulates focused F-DMAS beamforming starting from a PW DAS image without needing to access the raw RF data, which is typically unavailable with clinical ultrasound devices.
2022
978-1-6654-6657-8
File in questo prodotto:
File Dimensione Formato  
(2022) proceeding GAN BF IUS.pdf

non disponibili

Descrizione: published proceeding
Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Non Pubblico - Accesso privato/ristretto
Dimensione 2.7 MB
Formato Adobe PDF
2.7 MB Adobe PDF   Visualizza/Apri   Richiedi una copia
2117_submitted_version.pdf

accesso aperto

Descrizione: accepted manuscript
Tipologia: 2. Post-print / Author's Accepted Manuscript
Licenza: PUBBLICO - Tutti i diritti riservati
Dimensione 2.72 MB
Formato Adobe PDF
2.72 MB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/2983225