Cannula Localization Using Separate Plane Wave Ultrasound Measurements and a Deep Neural Network

Sonography is commonly used to monitor the insertion of a cannula into human tissue. However, good visibility in B-mode images is rarely guaranteed, especially with deep punctures and steep insertion angles. In previous work, we showed that resonance effects in the cannula can be observed in the received signals [1]. However, these nonlinear effects are not exploited in conventional ultrasound imaging. We hypothesize that the utilization of deep neural networks is particularly suitable to process this nonlinear cannula-related information. Moreover, we presume that the exploitation of different information found in different plane waves could enhance the localization of the cannula compared to the direct utilization of compounded B-mode images. Therefore, a deep learning-based framework capable of extracting the information provided by the separate plane waves is employed to enhance the localization of the shaft and tip of the cannula. The implemented architecture is based on a modified U-Net architecture. It predicts the position of the cannula from three different input scenarios: compounded B-mode images, separate plane wave B-mode images (3 angles), and separate plane wave B-mode images combined with their difference images. The model was first trained on data acquired in a water bath and then fine-tuned on a porcine dataset. This way, even with a comparatively small dataset, it was possible to achieve faster convergence and better localization results. Using separate plane wave B-mode images and their difference images substantially improved the visualization of the cannula by achieving 4.5-times lower tip localization error (measured by calculating the mean absolute difference between the true and the predicted tip location) and improving the estimation of the puncture angle by 30.6% (measured by calculating the mean absolute difference between the true and the predicted puncture angles) compared to the direct utilization of compounded B-mode images.