Quantification of thyroid nodules in multiple ultrasonography systems.
Authors
Affiliations (10)
Affiliations (10)
- School of Electrical Engineering Korea Advanced Institute of Science and Technology, Daejeon, South Korea. Electronic address: [email protected].
- Barreleye Inc., Seoul, South Korea. Electronic address: [email protected].
- Barreleye Inc., Seoul, South Korea. Electronic address: [email protected].
- School of Electrical Engineering Korea Advanced Institute of Science and Technology, Daejeon, South Korea. Electronic address: [email protected].
- School of Electrical Engineering Korea Advanced Institute of Science and Technology, Daejeon, South Korea. Electronic address: [email protected].
- School of Electrical Engineering Korea Advanced Institute of Science and Technology, Daejeon, South Korea. Electronic address: [email protected].
- School of Electrical Engineering Korea Advanced Institute of Science and Technology, Daejeon, South Korea. Electronic address: [email protected].
- Department of Emergency Medicine, Seoul National University Bundang Hospital, Seong-nam, South Korea. Electronic address: [email protected].
- Department of Radiology, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seong-nam, South Korea. Electronic address: [email protected].
- School of Electrical Engineering Korea Advanced Institute of Science and Technology, Daejeon, South Korea. Electronic address: [email protected].
Abstract
Quantitative ultrasound (QUS) has been proposed in recent studies to extract tissue acoustic properties from pulse-echo signal. This paper introduces QIT-net, a quantitative imaging technique designed to assess thyroid nodules by quantifying acoustic attenuation (ATT) and speed of sound for multiple ultrasonography systems. The proposed method employs a CNN-Transformer hybrid architecture to effectively capture local features for fine details and global features for macro context within RF data. B-mode images are employed as an auxiliary input to ensure reliable performance regardless of the complex structures present in the human neck. Additionally, a fine-tuning strategy by devices is incorporated to handle variations in B-mode image characteristics across different ultrasound devices. To train the deep neural model across diverse ultrasonography systems, we propose a network architecture that shares parameters across systems, enabling the model to learn common features from all available datasets. The proposed method is evaluated through numerical simulations, ex-vivo phantoms and clinical tests on both a research grade ultrasound device and a commercial ultrasound device.