Exhibit Hall | Forum 9
Purpose: Conventional 2D-ultrasound (US) imaging is an easy to use, affordable and non-ionizing modality, yet fails to offer a volumetric visualization of the organ. As a result, it is underutilized in many, but especially intraoperative and interventional applications, where repetitive 3D-imaging is required to facilitate safe resections. We suggest to eliminate this limitation by developing a new tomographic reconstruction method for freehand 2D-series, by incorporating noise decorrelation and transducer specific-waveforms for temporospatial alignment of the images. The development, generalizability and performance evaluation of this method is discussed in this work.
Methods: The proposed method takes advantage of the fact that scatter distributions correlation of two US images is affected by the distance and their orientation similarity with respect to the imaged object. Already this information will enable alignment of consecutive slices, yet its quality will be poor, since scatter is known to be a low resolution data. Therefore, we also utilize generalized transducer-type-specific waveform information to facilitate geometrically accurate alignment. The model is a 8-layer Convolutional Neural Network (CNN), which takes two given 2D-US images and associated waveforms as inputs, and outputs the expected Cartesian transform. CNN architecture consists of two series of 2x(5x5x64) convolutional and one 2x2 max pooling layer, followed by fully connected layers. It was trained using 70000 intraoperative US images of the liver (45 patients), recorder with electromagnetic racked transducers. Mean square error was used as a loss. Geometric accuracy of the method was evaluated using CIRS ATS-539 and Kyoto Kagaku US-3-IOSFAN ultrasound phantoms.
Results: The mean drift of the reconstructed volumes was 3mm, while the rotational accuracy remained below 2.5° (~clinically acceptable level). Stable performance for different transducers at all acquisition depth levels, except 40%, was confirmed.
Conclusion: a new tomographic reconstruction method for freehand ultrasound was developed and illustrates promising results for clinical implementation.
Funding Support, Disclosures, and Conflict of Interest: This project is supported by the AAPM Research Seed Grant 2021-2022