Acoustic angiography provides high-resolution ultrasound volumes of microvasculature with minimal tissue background. Based on superharmonic imaging, the technique leverages the enhanced response of nonlinear microbubbles relative to linear tissue at superharmonic frequencies. The high contrast-to-tissue ratio enables quantification of vascular features to differentiate between healthy and tumor-associated vasculature, but the method relies on time-intensive, cumbersome processing with limited utility in future clinical applications. A more computationally efficient end-to-end analysis is demonstrated using deep learning. Acoustic angiography volumes (n = 195) of tumors and controls were acquired in vivo in rodents, using a confocal dual-frequency transducer transmitting at 4 MHz and receiving at 30 MHz. Different convolutional neural networks were evaluated in a nested cross-validation study to simultaneously tune hyperparameters and compare model performance. The best performing model achieved a mean classification accuracy of 92.8 ± 3.4% and a mean area under the ROC curve score of 96.7 ± 1.8%, improving upon previously published results. Furthermore, model performance was validated against quantified vascular metrics, and significantly higher vascular tortuosity values were measured in high network attention regions in tumors compared to controls (p 0.05). These results establish the feasibility of deep learning for accurately and efficiently detecting cancer in acoustic angiography volumes.
Bautista et al. (Wed,) studied this question.