Deep Learning based Segmentation of Cervical Blood Vessels in Ultrasound Images

Background and goal of study: Puncture of central vessels is a
frequently used therapeutic and diagnostic procedure. The use
of ultrasound (US) during needle insertion has become the gold
standard. Handling the US probe and needle is challenging,
especially in dif ficult anatomic conditions. Our long-term vision is
a deep learning based and augmented reality (AR) assisted needle
puncture. We aim to visualize the vessel structures in 3D based on
2D US image segmentation. While punctuating, the relative needle
tip position and relevant vessels can be highlighted via AR lenses to
optimize the image guidance process.

Materials and methods: Our experimental setup (Fig. 1) allows
to record robot poses for 3D reconstruction1 and US images
simultaneously while moving the probe manually over the vessel
structures. We record a pre-clinical dataset consisting of 3445
US images of the v. jugularis and art. carotis from seven dif ferent
probands. The data is split into individual subsets for training and
testing of a neural network, LinkNet2, for segmentation.
Figure 1. Our experimental setup with the US probe mounted to the
robot (Panda, Franka Emika) and positioned at the v. jugularis with
a visualization of exemplar y US images along with the segmentation
label and the segmentation mask predicted by the neural network.
Results and discussion: We obtain the best segmentation results
for the LinkNet pretrained with a ResNet101 backbone, resulting in
a DICE score of 0.915 and a Jaccard Index of 0.847. The segmentation
masks of the vessels show a high amount of overlap to the
labels (Fig. 1) and capture the form of the vessels. Minor errors occur
in areas where the two vessels are too close in the underlying
US images.

Conclusion: Our results show that the LinkNet is capable of segmenting
the area of interest with high quality. It is a small network,
suf ficient for fast data processing. Future work can improve our results
using more data. Peripheral nerve block or puncture of groin
vessels are further possible applications, as well as training of US
inexperienced users.

References:
1. Virga, S., et al. “Automatic force-compliant robotic ultrasound
screening of abdominal aortic aneurysms.”IEEE/RSJ IROS, 2016.
2. Chaurasia, A., and E. Culurciello. “Linknet: Exploiting encoder
representations for ef ficient semantic segmentation.”IEEE VCIP,
2017.

Acknowledgements:
This work was partially funded by FMTHH
(grant: 03FMTHH20).