|M.Sc. candidate||Dany Granada|
|Thesis||Pose estimation for a robotically assisted steerable needle|
|Supervisors||Elena De Momi, Alberto Favaro, Francesco Amigoni|
|Description||The medical scene has seen an increasing interest on minimal invasive surgeries (MIS). One of several approaches is the use of steering probes which can be potentially employed for tumor biopsy, brachytherapy, deep brain stimulation, and localised drug delivery in soft tissues.
There are few alternatives for needle steering currently either in the market or investigated on the scientific scene. One of the most innovative approach was proposed by Secoli & Rodriguez y Baena, nominally the Programmable Bevel-tip Needle (PBN). This new solution has many clinical advantages, including the facility to reach regions in the anatomy not accessible by other practices.
One of question still to be solved is a precise and deployable control system for this device, mainly due to the complexity of the needle-tissue interaction, the three dimensional nature of the problem, the fact that several aspects of the needle are not observable from the exterior and the aggregated complexity due to the multiple inputs structure. This thesis project is focus on estimating the angle 𝜑, the roll angle, which has the effect of twisting the needle and heavily influencing the pose and the control action to drive the needle.
Our goal is to find a kinematic model of the torsion angle 𝜑, and to implement this model on a state observer (Uncented Kalman Filter), intended to incorporate new feedback information to the control in order to improve its performance.
|M.Sc. candidate||Gabriele Belotti|
|Thesis||Estimation of an end-effector pose under c-arm|
|Supervisors||Elena De Momi, Florent Nageotte, Bernard Bayle|
|Description||Initialization of an end effector pose is one of the key elements of a correct interventional radiology procedure.
If the end effector is carried by a robot-on-cart which is positioned just before a minimally invasive procedure we must provide the coordinates of the target in the global reference frame, the Operating Room, by means of stereoscopic images.
We would like to estimate it without additional sensors, with a sufficient precision to initialize an optimization problem (given the grade of precision required by a surgical procedure of this kind).
When the kuka iiwa (or any equivalent robot) is near the patient, making 3D acquisition is almost impossible, so we have to rely on projections to avoid collisions and prevent positioning issues, these are taken by a C-arm scanner.
This thesis has resumed the work of ICube lab on the development of a teleoperated needle grasper for interventional radiology for percutaneous applications.
The aim of this larger project is to reduce the exposure to radiation of surgical teams during interventional radiology procedures, using a robot to carry the needle.
The workflow included both the grasper development and a study of the possible trajectories of the needle (to avoid organs and access a specific point in 3D), while reducing the cost of the grasper by making it mostly 3D printed and by giving only 1 DoF to it.
|M.Sc. candidate||Sara Rustum|
|Thesis||Design, development and embedding of a linear actuator into the HydroJet, a soft-tethered endoscopic capsule that is maneuvered using water jet actuators, designed for UGI cancer screening in low and middle income countries.|
|Supervisors||Elena De Momi, Pietro Valdastri (UL)|
|Description||Gastric cancer is one of the most common malignancy in the world and the majority of cases occur in low-and middle-income countries (LMICs). Early detection of premalignant lesions through population based screening programs, decreases mortality. Because of the high initial cost of the equipment, the cost of the reprocessing of the devices and the difficulties in reaching rural areas, screening programs are limited in LMICs and remote locations.
This master thesis is part of a PhD thesis whose aims is the development of the HydroJet, a swallowable endoscopic capsule, provided with disposable and reusable components and designed to reduce the costs of the gastroscopy procedure and the reprocessing problem.
In particular, this thesis is focused on the design, development and embedding of a linear actuator into the HydroJet tether, aimed at introducing another degree of freedom for the enlargement of the workspace and at improving the manoeuvrability of the device.
|M.Sc. candidate||Alessandro Casella|
|Thesis||Fetal membrane segmentation for fetoscopic laser surgery|
|Supervisors||Elena De Momi, Leonardo S. Mattos (IIT), Dario Palladini, Sara Moccia (UNIVPM)|
|Description||Twin-to-twin transfusion syndrome (TTTS) is a complication of disproportionate blood supply, resulting in high fetus morbidity or mortality. Severe TTTS has more than 90% mortality rate. Laser therapy is among the main treatments for TTTS. The therapy involves endoscopic surgery using laser to interrupt the vessels that allow exchange of blood between fetuses. The placental membrane (Figure), which separates the two fetuses, is used as a reference by surgeons to detect the vessels that have to be treated. However membrane identification is a burdensome task also for expert surgeons due to issues related to (i) image quality (e.g., noise, varying illumination level, specular reflections), (ii) in-vivo image acquisition (e.g., limited field of view, different camera pose with respect to the membrane) and (iii) treated anatomical district (amniotic fluid, high patient variability).
While Convolutional Neural Networks (CNN) are widely used in segmentation tasks, there are no previous work concerning this pathology in literature. Due to the rareness of pathology, creating a dataset with a small number with a limited number of frames available introduces a further problem.On this background, the purpose of this thesis is to investigate new strategies to perform automatic segmentation and tracking in endoscopic videos using deep learning.This thesis is in collaboration with the Advanced Robotics Department (ADVR) at Istituto Italiano di Tecnologia (IIT) in Genoa (Italy).
|M.Sc. candidate||Emanuele Colleoni|
|Thesis||Deep Learning: 3D pose estimation of surgical instruments|
|Supervisors||Elena De Momi, Danail Stoyanov (UCL), Sara Moccia (UNIVPM)|
|Description||Pose estimation of the instruments in surgical videos refers to the act of estimating the coordinates of the joints of the instruments and the connections between them in order to generate a skeleton of the devices.
Automation of this task is a very important component for computer-assisted interventions, in particular in tele-robotic minimally invasive surgery, where the surgeon operates based on the images from the cameras of the robot.
While significant advantages have been made in recent years in this field, articulation detection is still a major challenge.
During last few years, many deep learning techniques have been used in order to perform instrument pose estimation starting from raw images from videos of the Da Vinci surgical robot (from MICCAI 2015 Challenge) and using 2D Fully Convolutional Neural Networks.
The weak point of these models is that they are not able to extract any temporal information from the inputs.The aim of this master thesis is to develop, in collaboration with the Robotic Vision Department at University College of London, a new neural network architecture using 3D convolutional layers in order to let the model to learn both spatial and temporal features directly from the videos or from clips of frames as shown in Figure and then to improve the accuracy of the pose estimation in comparison with the 2-dimensional models.
|M.Sc. candidate||Chiara Carlini|
|Thesis||Deep-learning based vessel tracking in minimally-invasive nephrectomy|
|Supervisors||Elena De Momi, Sara Moccia (UNIVPM)|
|Description||Vessel avoidance during nephrectomy is crucial to prevent bleeding and reduce after-treatment mortality or morbidity. In this vision, automatic vessel segmentation can be integrated in robot-assisted surgical procedures as to avoid robot tip to enter (forbidden) vascular regions. A strong literature already exists on vessel segmentation, however several limitations hamper the translation into the actual clinical practice, such as low robustness to illumination variation and noise, and to variability in patients and imaging instrumentation. Deep learning has the potential to overcome such limitations.
Thus, the goal of this thesis is to implement deep-learning based vessel segmentation and tracking in endoscopic videos recorded during nephrectomy. In particular, the implemented deep learning method is based on convolutional neural networks (CNNs). Many previous projects have obtained satisfying results in the detection and tracking of everyday life objects using CNNs starting from natural images captured by common webcams, as shown in Figure. For this reason, it has been decided to exploit CNNs potentiality to identify relevant anatomical structures present in real medical images, like endoscopic ones. In this specific case, CNNs are employed to segment kidney vessels from endoscopic kidney images. Some examples of the accomplished segmentation are shown in Figure.
|M.Sc. candidate||Ilaria Patrini|
|Thesis||Deep-learning models for informative-frame selection in laryngoscopic videos|
|Supervisors||Elena De Momi, Leonardo S. Mattos (IIT), Sara Moccia (UNIVPM)|
|Description||Early-stage diagnosis of laryngeal cancer is of primary importance to reduce patient morbidity. Narrow-band imaging (NBI) endoscopy is commonly used for screening purposes, reducing risks linked to a biopsy but at the cost of some drawbacks, such as large amount of data to review and/or analyze to make the diagnosis.
The purpose of this thesis is to investigate new strategies to perform automatic selection of informative endoscopic video frames (Figure), as to reduce the amount of data to process and potentially increase diagnosis performance. Several methods for informative-frame selection have been proposed in the literature, with drawbacks such as heavy parameter tuning and low robustness to inter- and intra-patient variability.
On this background, this thesis focuses on (1) investigating if the features learned with deep-learning models, such as Convolutional Neural Networks (CNNs), can outperform methods based on the handcrafted ones, (2) evaluate if CNNs for informative-frame selection over uninformative ones are more powerful than standard machine learning approaches, as to provide fast and accurate selection of informative video frames.
This thesis is in collaboration with the Advanced Robotics Department (ADVR) of Istituto Italiano di Tecnologia (IIT) in Genoa (Italy).
|M.Sc. candidate||Michela Ruperti|
|Thesis||Early-stage laryngeal-cancer diagnosis through learned-feature extraction and classification|
|Supervisors||Elena De Momi, Leonardo S. Mattos (IIT), Sara Moccia (UNIVPM)|
|Description||Early-stage diagnosis of laryngeal squamous cell carcinoma (SCC) is of primary importance for lowering patient mortality or after treatment morbidity. Early-stage SCC is associated with the presence of intrapapillary capillary loops (IPCL) and hypertrophic vessels (Hbv), as well as with the presence of pre-cancerous tissue conditions, such as leukoplakia (Le).
Despite the challenges in diagnosis reported in the clinical literature, few efforts have been invested in real time computer-assisted diagnosis.
The goal of this thesis is investigating deep-learning strategies to perform fast and accurate SCC diagnosis from endoscopic video-frames in narrow-band imaging (NBI). Examples of different laryngeal tissues the thesis is intended to analyze are shown in Figure. The work is in collaboration with the Department of Advanced Robotics @ Istituto Italiano di Tecnologia (IIT) in Genoa.
The thesis work exploits deep-learned features, instead of standard handcrafted ones, that are extracted using pre-trained Convolutional Neural Networks (CNNs), of which an example of their structure can be seen in Figure. The CNN models are pre-trained on the ImageNet dataset, a large scale database of natural images, and fine tuned to allow the classification of laryngeal images.
|M.Sc. candidate||Giulio Russo|
|Thesis||A handheld robotic tool (Micron) to automate vessel avoidance in microsurgery|
|Supervisors||Elena De Momi, Cameron Riviere (CMU), Sara Moccia (UNIVPM)|
|Description||In neurosurgery procedures for meningioma resections, minute and accurate operations in poor visibility conditions have to be performed. One of the most common complication is bleeding caused by tools misplacement with respect to the vessels.
Numerous robotic systems have been developed to assist surgeons in micro-neurosurgical interventions: this thesis investigates the use of Micron.Micron is a robotic handheld tool, it guarantees an automatic vessel avoidance thanks to several technologies:
– Position control guarantees tremor compensation and tracking position;
– Image analysis recognizes the vessels position;The aim of this thesis is to design a force control strategy: the interaction between tip and vessels cannot exceed certain force thresholds, just to avoid vessels damages and bleeding.A FBG force sensor (optical fiber sensor) is positioned on the tip Micron: it’s a custom 2DOF temperature-compensated force sensor developed at Johns Hopkins University (Figure). It has a 0.25 mN precision.Hence, in parallel to the position loop, a force loop is implemented (Figure). Only when the force thresholds are exceeded, the force loop yields modify the reference position of the position loop (the main force regulator is a PID).The control strategy is implemented in LabView and the final system will be tested on phantoms.
|M.Sc. candidate||Guido Caccianiga|
|Thesis||Training in robot-assisted surgery: impact of transcranial direct current stimulation (tDCS) on motor learning during complex tasks|
|Supervisors||Elena De Momi, Jeremy Brown (JHU), Andrea Mariani (SSSA)|
|Description||Considering the act of training a surgeon for Robotic Minimal Invasive Surgery (RMIS), this study focuses on the application of innovative technologies to enhance the performance of the trainee and hopefully deliver a more balanced and efficient learning curve. In particular, this work studies how delivering transcranial Direct Current Stimulation (tDCS) during the execution of a complex motor task (surgical robotic procedures) can measurably improve both the accuracy and the learning curve of the trainee. The project also addresses the carryover effect between training on a simulation platform and the actual performance on the clinical robot.
The first step will be to develop a surgical related task through which better exploit the motor learning in specific relation to an error signal acting as a real time feedback to the user. A double blind randomized trial is set up to evaluate the effectiveness of the stimulation and the motor tasks are performed both on real and virtual environments. In vitro experiments are performed with the clinical da Vinci Surgical System by Intuitive Surgical, a minimally-invasive surgery robot. Virtual Reality experiments are performed with the dVRK which is the Research Kit for the da Vinci Surgical System. The replication of the same experiments on different platforms acts as a validation process for the VR training environment, and aims at the evaluation of how the skills acquired with the simulator are carried on the real surgical platform.
|M.Sc. candidate||Francesco Piquè|
|Thesis||Dynamic Modeling of the Da Vinci Research Kit Arm for the Estimation of Interaction Wrench|
|Supervisors||Elena De Momi, Arianna Menciassi (SSSA)|
|Description||The commercialized version of the da Vinci robot currently lacks of the haptic feedback to the master arm. Thus, the surgeon has no haptic sense and relies only on the visual feedback. For this reason, in the recent research activities using the da Vinci Research Kit (dVRK) platform, the reflection of the interaction force between the slave tool and the environment to the master manipulator is a topic of high interest.
The aim of this thesis is to implement and validate a sensorless model-based approach for contact force and torque estimation. The dynamic model of the dVRK slave arm is obtained and its parameters are identified. The idea is to use the joint torques obtained from the measured motor currents and to subtract the torques resulting from the dynamics of the robot arm. The resulting torques are therefore only due to the external forces and torques acting on the tool, which are then obtained through the inverse transpose of the Jacobian matrix. The accuracy of this method is assessed by comparing the estimated wrench to the one measured by a force/torque sensor (ATI mini 45).
|M.Sc. candidate||Mattia Pesenti|
|Thesis||Control-oriented modeling of the EMG-Force relationship to enhance Human-Robot Interaction|
|Supervisors||Elena De Momi, Bernard Bayle|
|Description||Characterizing the behavior of human operators has become a very relevant subject with the increasing popularity of human-robot interaction (HRI). This problem is addressed in the present master thesis project by modeling human operators during collaborative manipulation or telemanipulation tasks. Studying the dynamics of interaction on the human side could allow mimicking his behavior and/or compensate for it on the robot side. A primary objective is to improve the quality of robot-assisted surgery. In this field, it is of utmost importance to ensure stability of the interaction between the surgeon and the robot, but in the same time it also very important to offer the best transparency possible. In terms of control theory, it is to be related with the performance of control laws, an objective that is very well known to be conflicting with stability.
A novel modeling approach is considered to identify the EMG-Force relationship of the human arm. Such a model should allow to easily estimate the force developed by the human user during interaction simply measuring the EMG signals on the arm. Furthermore, it should be able to be integrated in the control scheme of a collaborative robot, in order to improve the interaction with the human.
|M.Sc. Candidate||Stefano De Nigris|
|Advisor||Elena De Momi, Digna M González-Otero (EHU-UPV), Jesús Ruiz(EHU-UPV), Sofía Ruiz de Gauna(EHU-UPV)|
|Title||Application of accelerometer-based chest compression feedback devices in novel scenarios|
|Collaborations||Laboratorio GSC (Grupo Señales y Comunicaciones), Universidad del País Vasco (EHU-UPV), Bilbao, Spain|
Sudden cardiac arrest is the largest cause of natural death worldwide. Early cardiopulmonary resuscitation (CPR) is key for patient survival. Resuscitation guidelines emphasize the importance of providing high quality chest compressions, that is, with a rate of 100 compressions per minute and a depth of 5cm in adults and one third of the diameter of the chest in children. However, meeting these requirements is difficult, even for well-trained rescuers.
The use of feedback systems can help rescuers increase chest compression quality. Most of these devices measure the acceleration of the chest during compressions, and guide the rescuer towards the target rate and depth. However, these systems present two main limitations. First, they are designed for adult patients, and thus present a target depth of 5cm, which is inadequate for children. Second, they could be inaccurate when used in moving vehicles, such as a train or a plane, as they would register the acceleration of the vehicle along with that of the patient’s chest.
The work tackles those two limitations. It is divided in two main objectives: First, to develop and test an algorithm to measure the diameter of the chest using an accelerometer. This algorithm could be used to adapt feedback systems to be used in pediatric patients. Second, to propose and test an additive model to evaluate the accuracy of accelerometer-based CPR feedback devices in moving vehicles, and to apply it to the case of a plane and train.
The results of this thesis could extend the scenarios of application of feedback systems, allowing their use in pediatric cardiac arrests or in public transportation means.