Feasibility of deep learning-based cancer detection in ultrasound microvascular images

Acoustic angiography is a superharmonic contrast-enhanced ultrasound modality that maps 3-D microvasculature with fine spatial resolutions and has demonstrated potential to improve disease detection. However, the application of acoustic angiography for cancer detection currently faces challenges. Quantitative analysis relies on time-consuming, manual segmentation of individual vessels, and inter-operator variability limits reader-based discrimination. This feasibility study aims to address the limitations of current approaches with deep learning for efficient and accurate detection of tumor-associated vasculature in vivo and to validate against quantitative methods that evaluate vascular morphology. Convolutional neural networks (CNNs), namely EfficientNet, ResNet, and DenseNet, were trained on a newly collected dataset of acoustic angiography volumes (n = 195 with 98 controls and 97 tumors) in rodents using a nested cross-validation study. The best performing model, 3-D EfficientNet-B0, achieved a mean classification accuracy of 0.928 ± 0.034 with high sensitivity and specificity, comparable to previously published results. Comparison with quantitative methods in tumor cases showed correlation between high network attention regions and morphological features typically associated with malignant vessels, including increased density and tortuosity. These results highlight the efficiency and accuracy of end-to-end CNNs for tumor detection in acoustic angiography volumes, validated by known markers of malignancy.