CUCT-Net: End-to-End Signal-to-Image Learning for Quantized Speed-of-Sound Estimation and Tissue Segmentation in Ultrasound Computed Tomography.

objectiveTraditional Full Waveform Inversion (FWI) methods for Ultrasound Computed Tomography (UCT) are computationally expensive and can be sensitive to strong acoustic contrasts. In this work, we propose the Multi-Channel Transducer Network (CUCT-Net), a deep learning framework that directly maps received ultrasound signals to image-space outputs for quantized speed-of-sound (SoS) estimation and for direct tissue-level segmentation over both low- and high-contrast regions, enabling end-to-end recovery of both contrast-driven and anatomically meaningful structures from raw measurements.

methodCUCT-Net uses a multi-input encoder-decoder architecture that maps raw multi-static UCT measurements to quantized SoS (or tissue-class) maps without requiring an initial guess or iterative optimization. Parallel per-transducer encoders extract view-specific features that are fused and refined by a decoder, with Shift Units (SU) used to enhance fine-scale feature modeling under sparse sensing. Experiments are performed on k-Wave simulations using (i) Shepp-Logan-inspired disc phantoms with Original/Distorted/Mixed variants and (ii) DBB-derived anatomical brain phantoms, under clean and noisy measurement conditions.

resultsThe proposed network achieves accurate quantized SoS estimation and direct tissue-level segmentation across synthetic and anatomically derived phantom experiments. Strong robustness to noise is demonstrated through transfer learning. Compared with FWI, CUCT-Net significantly reduces computational cost while maintaining stable performance under reduced-sensor conditions for quantized SoS estimation and complex tissue heterogeneity for segmentation.

conclusionsCUCT-Net formulates UCT as a direct signal-to-image learning problem that supports both quantized SoS estimation and tissue-level segmentation. By learning an end-to-end mapping from raw ultrasound measurements to quantized SoS or tissue representations, the proposed framework bypasses iterative inversion and achieves efficient and robust performance under reduced-sensor and strong-contrast conditions. The multi-input architecture enables effective integration of information from multiple transducers, demonstrating the feasibility and potential of data-driven end-to-end quantized SoS estimation and tissue segmentation for UCT.