Exploiting Resonance Effects for Cannula Localization using Multiview Spectral Channel Data as Input for a Deep Neural Network (DNN)

Sonography is a frequently employed method for monitoring cannula insertion into human tissue. Nonetheless, ensuring optimal visibility is often challenging, particularly in cases involving deep punctures and steep puncture angles. In previous work, it was shown that resonance effects in the cannula can be observed in the received signals. Later, we demonstrated the enhancements achieved in cannula localization by leveraging the diverse resonance effects observed in the B-mode images of the separate plane waves compared to the compounded images through a deep-learning based framework. However, we hypothesize that the results can be further improved by exploiting the resonance effects found in the raw radiofrequency (RF) data instead of relying solely on B-mode images. Therefore, a two-stage deep learning-based framework capable of extracting cannula-related information from the short time Fourier Transform of the raw RF data across separate plane waves is employed. The implemented architecture is based on a DenseNet backbone. In a first step, the network is trained on all steering angles separately. This is followed by a multi-view step, wherein parallel paths are incorporated, each encompassing a distinct steering angle. The localization results are compared to those achieved by the beamformed compounded (bc) input for reference. The model was first trained on data acquired in a water bath and then fine-tuned on a dataset acquired in porcine tissue. Utilizing the multi-view network with separate plane wave RF data, as opposed to using bc RF data, substantially reduced the mean absolute axial and lateral tip localization errors from 2.94 mm and 2.31 mm to 0.97 mm and 0.89 mm, respectively. Moreover, the proposed approach improved the F1 score of the puncture angle estimation from 96% to 100%.