Multitask affinity learning for quantitative ultrasound imaging

Recent advances in deep learning have brought great progress in quantitative tissue imaging. The performance of the neural network is highly dependent on the composition of the trained dataset. However, it is rarely possible to recognize the actual biomechanical characteristics of the observed human tissue, hence constructing high-quality quantitative ultrasound data became a crucial challenge. In this paper, we propose an affinity learning framework that utilizes the pixel-wise resemblance of the b-mode image for generating the quantitative ultrasound image. The proposed framework contributes to the reconstruction accuracy of the quantitative ultrasound images by training high-fidelity in-vivo ultrasound signals without annotation of tissue bio-mechanical value. The proposed method was evaluated through numerical simulations, phantom, and in-vivo measurements. In in-vivo studies, the proposed neural network reconstructs detailed quantitative images and demonstrates clinical potential in the differential diagnosis of breast cancer.