Section: New Results

Medical robotics

Visual servoing based on ultrasound images

Participants : Caroline Nadeau, Alexandre Krupa.

We developed a new approach of ultrasound image based visual servoing that directly uses the intensities of the ultrasound image pixels as visual features. This method that spares any segmentation or image processing time consuming step was initially proposed to control the 6 DOF of a conventional 2D probe for positioning and tracking tasks [38] , [48] . To increase the tracking performance we also adapted this method by considering a predictive control law based on the periodicity of physiological motions [36] . Rigid motion compensation experiments were conducted in the context of the ANR USComp project (See Section  8.2.3 ). The method was also improved by estimating on-line the image 3D gradient required for the positioning task and extended for the use of a bi-plane ultrasound probe [37] . Finally, the use of a 3D motorized probe was also considered to compute directly the image 3D gradient and a comparison of the results obtained with the different probes (2D, bi-plan, 3D) was performed [12] .

Autonomous control modes for ultrasound probe guidance

Participants : Tao Li, Alexandre Krupa.

In the context of the ANR Prosit (See Section  8.2.2 ), we developed several autonomous control modes in order to assist a doctor during a robotized and teleoperated ultrasound examination (tele-echography). The robotic tasks we proposed concern: an automatic scanning of the patient by a 2D probe, a shared control mode that maintains the visibility of an anatomic element of interest while the doctor teleoperates the slave robot holding the 2D probe, an automatic positioning task that allows the doctor to retrieve a desired anatomic section that was previously captured by the doctor. The two latter modes are based on visual servoing schemes that use as input image moments extracted from the observed 2D ultrasound image. This extraction is performed thanks to an active contour (snake) based on Fourier descriptors that we developed and implemented on GPU in order to provide real-time performance [34] , [47] . The proposed autonomous control modes were experimentally validated on the Lagadic medical robotics platform (see Section  5.3 ) and are now in the process of being integrated on the Prosit robot platform.

Real-time 3D ultrasound image reconstruction and 3D deformation tracking

Participants : Deukhee Lee, Alexandre Krupa.

We developed and implemented on GPU an algorithm that reconstructs in real-time a sequence of dense ultrasound volumes from a set of pre-scan 2D ultrasound images provided online by a motorized ultrasound probe [33] . Then we proposed a dense ultrasound tracking algorithm that estimates in real time both rigid and non-rigid motions of a region of interest observed in the sequence of reconstructed ultrasound volumes [33] . The algorithm consists in estimating in real-time, from intensity-value changes between successive 3D ultrasound images, motions of a set of 3D control points that describe the evolution of 3D Thin-Plate Splines (TPS) modeling the deformation. The estimated rigid motion was then used in a pose-based control scheme to automatically displace the probe held by a robot for soft tissue motion compensation. These works were conducted in the context of the ANR USComp project (See Section  8.2.3 ).