News & Events

An Investigation into the Use of Machine Learning Techniques to Enhance the Performance of Ultrasound and Photoacoustic Imaging

Bo Wen

Supervised by Marvin Doyley

703 Computer Studies Building

Ultrasound and photoacoustic imaging provide complementary information about tissue structure and physiology, but the quality of the recon-structed image is closely tied to the way the data are acquired. In plane-wave ultrasound, reducing the number of transmissions increases temporal resolution but sacrifices image quality. Coherent compounding can recover much of this loss by combining data from multiple steering angles, although this advantage comes at the cost of a longer acquisition and relies on the assumption that the imaged tissue remains stationary during the angular sequence. This assumption becomes increasingly difficult to satisfy in cardiac imaging. Photoacoustic imaging presents a different limitation. Weak signals, depth-dependent attenuation, and experimental noise can obscure vascular structures, while a noise-free in vivo reference image is generally not available for supervised training.

The objective of this thesis is to investigate how learning-based image reconstruction can be used under these different acquisition and supervision constraints. Particular attention is given to the information contained in the training target, because the reference used to supervise a network is itself a product of the imaging experiment and is not necessarily a complete or physically exact representation of the desired image.

This question was initially studied in single-plane-wave ultrasound imaging. Raw RF channel data from one unsteered transmission were used to re-construct an image referenced to a 33-angle coherent plane-wave compounded acquisition. A Transformer-based channel-domain model improved image quality relative to the single-angle input in simulation and physical phantom experiments and approached the compounded reference in several image-quality measures. The reconstructed images nevertheless contained struc-tured errors in some simulated targets even when the aggregate image-quality metrics were favorable. This result motivated the introduction of anatomical information into the reconstruction problem. A hybrid U-Net/Transformer framework was subsequently developed for joint reconstruction and segmentation, with structural information supplied through known target geometry in simulation and phantom experiments and through manually defined carotid-lumen masks in the in vivo study. When the network configuration was held fixed and the training target was altered, the structurally informed target was associated with improved lumen delineation. The evaluated in vivo configurations achieved median Dice scores of approximately 0.92 while maintaining reconstruction performance comparable to the compounded reference.

The use of a compounded image as a supervisory reference becomes less straightforward when the anatomy moves during acquisition. To study this problem under controlled conditions, a dynamic cardiac ultrasound simulation was constructed from ACDC-derived anatomy and Field II channel-data generation. For each cardiac state, the scatterer distribution was held fixed while the single-angle input and an eleven-angle coherent plane-wave com-pounded reference were generated. This frozen-phase construction removed inter-angle motion from the reference within the simulation and allowed an-gular compounding to be examined separately from cardiac motion. An attention-based ResUNet reconstructed the large-scale cardiac structure and blood–myocardium contrast of the phase-matched reference from a single simulated transmission. The reconstructed images showed improved blood–myocardium separability relative to single-angle delay-and-sum imaging, but they were consistently smoother than the frozen-phase reference. These experiments establish feasibility within the simulated acquisition and motion model rather than experimental cardiac performance.

A different form of supervision was required for volumetric photoacoustic foot imaging, where a clean in vivo target could not be measured directly. Training pairs were therefore constructed from experimentally acquired data by using the original acquisition as a light-noise reference and reinjecting amplified measured noise to form a more strongly degraded input. Because adjacent elevational slices contain related vascular structure, a bidirectional ConvLSTM was used to incorporate information from slices on both sides of the frames being restored. The method was evaluated with paired in vivo data, a physical vessel phantom, and independent reference-free acquisitions. It produced higher SNR and lower reconstruction error than the learned baselines in the paired evaluation, showed vessel-width agreement closest to the known phantom geometry, and increased image-based vessel-continuity measures in the reference-free data.

The studies in this thesis show that the interpretation of a learned re-construction depends on more than the network used to generate it. The available measurements determine what information can be reconstructed, the supervisory reference determines what information the model is encouraged to reproduce, and the evaluation criteria determine which aspects of the reconstruction are considered successful. These factors must therefore be considered together when learning-based reconstruction is applied to imaging problems in which acquisition or reference data are inherently constrained.