Ultrasound (US) imaging stands as a valid alternative to X-rays based methodologies for vascular screening and intraoperative navigation. However, US images quality is highly operator dependent, thus to standardize image acquisition routine and increase interpretability Robotic US Systems (RUSS) and deep learning strategies have been recently developed. Artificial intelligence applications typically consist in automatic vessels segmentation method, however, in intraoperative scenarios it’s impossible to guarantee that every frame contains a vessel to be segmented. Thus, to increase robustness of such applications, we propose a multi-task convolutional neural network (CNN) architecture able to distinguish between vessel and no vessel images, in addition to segmenting them.

The architecture is a modified version of U-Net, obtained by adding a classification branch, after the encoder portion, that is able to detect the presence of vessels in the image. Transfer learning was adopted, enabling the architecture training with a relatively small size dataset (i.e., 240 images). The average classification accuracy and segmentation dice similarity coefficient were both higher than 90% over a 6-fold cross-validation with a computation time lower than 10 ms per image.

These preliminary results indicate that such multi-task CNN could be efficiently integrated in a robotic platform, potentially enabling robust visual-servoing procedures. Additionally, application can be enlarge to different districts by further fine-tuning the network exploiting transfer learning and small datasets.