Evaluating the impact of reinforcement learning on automatic deep brain stimulation planning.

PURPOSE: Traditional techniques for automating the planning of brain electrode placement based on multi-objective optimization involving many parameters are subject to limitations, especially in terms of sensitivity to local optima, and tend to be replaced by machine learning approaches. This paper explores the feasibility of using deep reinforcement learning (DRL) in this context, starting with the single-electrode use-case of deep brain stimulation (DBS). METHODS: We propose a DRL approach based on deep Q-learning where the states represent the electrode trajectory and associated information, and actions are the possible motions. Deep neural networks allow to navigate the complex state space derived from MRI data. The chosen reward function emphasizes safety and accuracy in reaching the target structure. The results were compared with a reference (segmented electrode) and a conventional technique. RESULTS: The DRL approach excelled in navigating the complex anatomy, consistently providing safer and more precise electrode placements than the reference. Compared to conventional techniques, it showed an improvement in accuracy of 2.3% in average proximity to obstacles and 19.4% in average orientation angle. Expectedly, computation times rose significantly, from 2 to 18 min. CONCLUSION: Our investigation into DRL for DBS electrode trajectory planning has showcased its promising potential. Despite only delivering modest accuracy gains compared to traditional methods in the single-electrode case, its relevance for problems with high-dimensional state and action spaces and its resilience against local optima highlight its promising role for complex scenarios. This preliminary study constitutes a first step toward the more challenging problem of multiple-electrodes planning.

Generation of synthetic training data for SEEG electrodes segmentation.

PURPOSE: Stereoelectroencephalography (SEEG) is a minimally invasive surgical procedure, used to locate epileptogenic zones. An accurate identification of the metallic contacts recording the SEEG signal is crucial to ensure effectiveness of the upcoming treatment. However, due to the presence of metal, post-operative CT scans contain strong streak artefacts that interfere with deep learning segmentation algorithms and require a lot of training data to distinguish from actual contacts. We propose a method to generate synthetic data and use them to train a neural network to precisely locate SEEG electrode contacts. METHODS: Random electrodes were generated following manufacturer's specifications and dimensions and placed in acceptable regions inside metal-free CT images. Metal artefacts were simulated in the generated data set using radon transform, beam hardening, and filtered back projection. A UNet neural network was trained for the contacts segmentation task using various training set-ups combining real data, basic augmented data, and synthetic data. The results were compared. RESULTS: We reported a higher accuracy when including synthetic data during the network training, while training only on real and basic augmented data more often led to misclassified artefacts or missed contacts. The network segments post-operative CT slices in less than 2 s using 4 GeForce RTX2080 Ti GPUs and in under a minute using a standard PC with GeForce GTX1060. CONCLUSION: Using synthetic data to train the network significantly improves contact detection and segmentation accuracy.

A self-supervised learning strategy for postoperative brain cavity segmentation simulating resections.

PURPOSE: Accurate segmentation of brain resection cavities (RCs) aids in postoperative analysis and determining follow-up treatment. Convolutional neural networks (CNNs) are the state-of-the-art image segmentation technique, but require large annotated datasets for training. Annotation of 3D medical images is time-consuming, requires highly trained raters and may suffer from high inter-rater variability. Self-supervised learning strategies can leverage unlabeled data for training. METHODS: We developed an algorithm to simulate resections from preoperative magnetic resonance images (MRIs). We performed self-supervised training of a 3D CNN for RC segmentation using our simulation method. We curated EPISURG, a dataset comprising 430 postoperative and 268 preoperative MRIs from 430 refractory epilepsy patients who underwent resective neurosurgery. We fine-tuned our model on three small annotated datasets from different institutions and on the annotated images in EPISURG, comprising 20, 33, 19 and 133 subjects. RESULTS: The model trained on data with simulated resections obtained median (interquartile range) Dice score coefficients (DSCs) of 81.7 (16.4), 82.4 (36.4), 74.9 (24.2) and 80.5 (18.7) for each of the four datasets. After fine-tuning, DSCs were 89.2 (13.3), 84.1 (19.8), 80.2 (20.1) and 85.2 (10.8). For comparison, inter-rater agreement between human annotators from our previous study was 84.0 (9.9). CONCLUSION: We present a self-supervised learning strategy for 3D CNNs using simulated RCs to accurately segment real RCs on postoperative MRI. Our method generalizes well to data from different institutions, pathologies and modalities. Source code, segmentation models and the EPISURG dataset are available at https://github.com/fepegar/resseg-ijcars .

Surgical planning assistance in keyhole and percutaneous surgery: A systematic review.

Surgical planning of percutaneous interventions has a crucial role to guarantee the success of minimally invasive surgeries. In the last decades, many methods have been proposed to reduce clinician work load related to the planning phase and to augment the information used in the definition of the optimal trajectory. In this survey, we include 113 articles related to computer assisted planning (CAP) methods and validations obtained from a systematic search on three databases. First, a general formulation of the problem is presented, independently from the surgical field involved, and the key steps involved in the development of a CAP solution are detailed. Secondly, we categorized the articles based on the main surgical applications, which have been object of study and we categorize them based on the type of assistance provided to the end-user.

Dynamic path planning for percutaneous procedures in the abdomen during free breathing.

PURPOSE: Percutaneous procedures are increasingly used for the treatment of tumors in abdominal structures. Most of the time, these procedures are planned based on static preoperative images and do not take into account any motions, while breathing control is not always applicable. In this paper, we present a method to automatically adjust the planned path in real time according to the breathing. METHODS: First, an estimation of the organs motions during breathing is performed during an observation phase. Then we propose an approach named Real Time Intelligent Trajectory (RTIT) that consists in finding the most appropriate moments to push the needle along the initially planned path, based on the motions and the distance to surrounding organs. We also propose a second approach called Real Time Straight Trajectory (RTST) that examines sixteen scenarios of needle insertion at constant speed, starting at eight different moments of the breathing cycle with two different speeds. RESULTS: We evaluated our methods on six 3D models of abdominal structures built using image datasets and a real-time simulation of breathing movements. We measured the deviation from the initial path, the target positioning error, and the distance of the actual path to risky structures. The path proposed by RTIT approach is compared to the best path proposed by RTST. CONCLUSIONS: We show that the RTIT approach is relevant and adapted to breathing movements. The modification of the path remains minimal while collisions with obstacles are avoided. This study on simulations constitutes a first step towards intelligent robotic insertion under real-time image guidance.

GPU-based 3D iceball modeling for fast cryoablation simulation and planning.

PURPOSE: The elimination of abdominal tumors by percutaneous cryoablation has been shown to be an effective and less invasive alternative to open surgery. Cryoablation destroys malignant cells by freezing them with one or more cryoprobes inserted into the tumor through the skin. Alternating cycles of freezing and thawing produce an enveloping iceball that causes the tumor necrosis. Planning such a procedure is difficult and time-consuming, as it is necessary to plan the number and cryoprobe locations and predict the iceball shape which is also influenced by the presence of heating sources, e.g., major blood vessels and warm saline solution, injected to protect surrounding structures from the cold. METHODS: This paper describes a method for fast GPU-based iceball modeling based on the simulation of thermal propagation in the tissue. Our algorithm solves the heat equation within a cube around the cryoprobes tips and accounts for the presence of heating sources around the iceball. RESULTS: Experimental results of two studies have been obtained: an ex vivo warm gel setup and simulation on five retrospective patient cases of kidney tumors cryoablation with various levels of complexity of the vascular structure and warm saline solution around the tumor tissue. The experiments have been conducted in various conditions of cube size and algorithm implementations. Results show that it is possible to obtain an accurate result within seconds. CONCLUSION: The promising results indicate that our method yields accurate iceball shape predictions in a short time and is suitable for surgical planning.

Comparison of interactive and automatic segmentation of stereoelectroencephalography electrodes on computed tomography post-operative images: preliminary results.

Stereoelectroencephalography is a surgical procedure used in the treatment of pharmacoresistant epilepsy. Multiple electrodes are inserted in the patient's brain in order to record the electrical activity and detect the epileptogenic zone at the source of the seizures. An accurate localisation of their contacts on post-operative images is a crucial step to interpret the recorded signals and achieve a successful resection afterwards. In this Letter, the authors propose interactive and automatic methods to help the surgeon with the segmentation of the electrodes and their contacts. Then, they present a preliminary comparison of the methods in terms of accuracy and processing time through experimental measurements performed by two users, and discuss these first results. The final purpose of this work is to assist the neurosurgeons and neurologists in the contacts localisation procedure, make it faster, more precise and less tedious.

Automatic planning of needle placement for robot-assisted percutaneous procedures.

PURPOSE: Percutaneous procedures allow interventional radiologists to perform diagnoses or treatments guided by an imaging device, typically a computed tomography (CT) scanner with a high spatial resolution. To reduce exposure to radiations and improve accuracy, robotic assistance to needle insertion is considered in the case of X-ray guided procedures. We introduce a planning algorithm that computes a needle placement compatible with both the patient's anatomy and the accessibility of the robot within the scanner gantry. METHODS: Our preoperative planning approach is based on inverse kinematics, fast collision detection, and bidirectional rapidly exploring random trees coupled with an efficient strategy of node addition. The algorithm computes the allowed needle entry zones over the patient's skin (accessibility map) from 3D models of the patient's anatomy, the environment (CT, bed), and the robot. The result includes the admissible robot joint path to target the prescribed internal point, through the entry point. A retrospective study was performed on 16 patients datasets in different conditions: without robot (WR) and with the robot on the left or the right side of the bed (RL/RR). RESULTS: We provide an accessibility map ensuring a collision-free path of the robot and allowing for a needle placement compatible with the patient's anatomy. The result is obtained in an average time of about 1 min, even in difficult cases. The accessibility maps of RL and RR covered about a half of the surface of WR map in average, which offers a variety of options to insert the needle with the robot. We also measured the average distance between the needle and major obstacles such as the vessels and found that RL and RR produced needle placements almost as safe as WR. CONCLUSION: The introduced planning method helped us prove that it is possible to use such a "general purpose" redundant manipulator equipped with a dedicated tool to perform percutaneous interventions in cluttered spaces like a CT gantry.

Self-guided training for deep brain stimulation planning using objective assessment.

OBJECTIVE: Deep brain stimulation (DBS) is an increasingly common treatment for neurodegenerative diseases. Neurosurgeons must have thorough procedural, anatomical, and functional knowledge to plan electrode trajectories and thus ensure treatment efficacy and patient safety. Developing this knowledge requires extensive training. We propose a training approach with objective assessment of neurosurgeon proficiency in DBS planning. METHODS: To assess proficiency, we propose analyzing both the viability of the planned trajectory and the manner in which the operator arrived at the trajectory. To improve understanding, we suggest a self-guided training course for DBS planning using real-time feedback. To validate the proposed measures of proficiency and training course, two experts and six novices followed the training course, and we monitored their proficiency measures throughout. RESULTS: At baseline, experts planned higher quality trajectories and did so more efficiently. As novices progressed through the training course, their proficiency measures increased significantly, trending toward expert measures. CONCLUSION: We developed and validated measures which reliably discriminate proficiency levels. These measures are integrated into a training course, which quantitatively improves trainee performance. The proposed training course can be used to improve trainees' proficiency, and the quantitative measures allow trainees' progress to be monitored.

Automatic preoperative planning of DBS electrode placement using anatomo-clinical atlases and volume of tissue activated.

PURPOSE: Deep brain stimulation (DBS) is a procedure requiring accurate targeting and electrode placement. The two key elements for successful planning are preserving patient safety by ensuring a safe trajectory and creating treatment efficacy through optimal selection of the stimulation point. In this work, we present the first approach of computer-assisted preoperative DBS planning to automatically optimize both the safety of the electrode's trajectory and location of the stimulation point so as to provide the best clinical outcome. METHODS: Building upon the findings of previous works focused on electrode trajectory, we added a set of constraints guiding the choice of stimulation point. These took into account retrospective data represented by anatomo-clinical atlases and intersections between the stimulation region and sensitive anatomical structures causing side effects. We implemented our method into automatic preoperative planning software to assess if the algorithm was able to simultaneously optimize electrode trajectory and the stimulation point. RESULTS: Leave-one-out cross-validation on a dataset of 18 cases demonstrated an improvement in the expected outcome when using the new constraints. The distance to critical structures was not reduced. The intersection between the stimulation region and structures sensitive to stimulation was minimized. CONCLUSIONS: Introducing these new constraints guided the planning to select locations showing a trend toward symptom improvement, while minimizing the risks of side effects, and there was no cost in terms of trajectory safety.

Preoperative trajectory planning for percutaneous procedures in deformable environments.

In image-guided percutaneous interventions, a precise planning of the needle path is a key factor to a successful intervention. In this paper we propose a novel method for computing a patient-specific optimal path for such interventions, accounting for both the deformation of the needle and soft tissues due to the insertion of the needle in the body. To achieve this objective, we propose an optimization method for estimating preoperatively a curved trajectory allowing to reach a target even in the case of tissue motion and needle bending. Needle insertions are simulated and regarded as evaluations of the objective function by the iterative planning process. In order to test the planning algorithm, it is coupled with a fast needle insertion simulation involving a flexible needle model and soft tissue finite element modeling, and experimented on the use-case of thermal ablation of liver tumors. Our algorithm has been successfully tested on twelve datasets of patient-specific geometries. Fast convergence to the actual optimal solution has been shown. This method is designed to be adapted to a wide range of percutaneous interventions.

Statistical study of parameters for deep brain stimulation automatic preoperative planning of electrodes trajectories.

PURPOSE: Automatic methods for preoperative trajectory planning of electrodes in deep brain stimulation are usually based on the search for a path that resolves a set of surgical constraints to propose an optimal trajectory. The relative importance of each surgical constraint is usually defined as weighting parameters that are empirically set beforehand. The objective of this paper is to analyze the use of these parameters thanks to a retrospective study of trajectories manually planned by neurosurgeons. For that purpose, we firstly retrieved weighting factors allowing to match neurosurgeons manually planned choice of trajectory on each retrospective case; secondly, we compared the results from two different hospitals to evaluate their similarity; and thirdly, we compared the trends to the weighting factors empirically set in most current approaches. METHODS: To retrieve the weighting factors best matching the neurosurgeons manual plannings, we proposed two approaches: one based on a stochastic sampling of the parameters and the other on an exhaustive search. In each case, we obtained a sample of combinations of weighting parameters with a measure of their quality, i.e., the similarity between the automatic trajectory they lead to and the one manually planned by the surgeon as a reference. Visual and statistical analyses were performed on the number of occurrences and on the rank means. RESULTS: We performed our study on 56 retrospective cases from two different hospitals. We could observe a trend of the occurrence of each weight on the number of occurrences. We also proved that each weight had a significant influence on the ranking. Additionally, we observed no influence of the medical center parameters, suggesting that the trends were comparable in both hospitals. Finally, the obtained trends were confronted to the usual weights chosen by the community, showing some common points but also some discrepancies. CONCLUSION: The results tend to show a predominance of the choice of a trajectory close to a standard direction. Secondly, the avoidance of the vessels or sulci seems to be sought in the surroundings of the standard position. The avoidance of the ventricles seems to be less predominant, but this could be due to the already reasonable distance between the standard direction and the ventricles. The similarity of results between two medical centers tends to show that it is not an exceptional practice. These results suggest that manual planning software may introduce a bias in the planning by proposing a standard position.

Anticipation of brain shift in Deep Brain Stimulation automatic planning.

Deep Brain Stimulation is a neurosurgery procedure consisting in implanting an electrode in a deep structure of the brain. This intervention requires a preoperative planning phase, with a millimetric accuracy, in which surgeons decide the best placement of the electrode depending on a set of surgical rules. However, brain tissues may deform during the surgery because of the brain shift phenomenon, leading the electrode to mistake the target, or moreover to damage a vital anatomical structure. In this paper, we present a patient-specific automatic planning approach for DBS procedures which accounts for brain deformation. Our approach couples an optimization algorithm with FEM based brain shift simulation. The system was tested successfully on a patient-specific 3D model, and was compared to a planning without considering brain shift. The obtained results point out the importance of performing planning in dynamic conditions.

PyDBS: an automated image processing workflow for deep brain stimulation surgery.

PURPOSE: Deep brain stimulation (DBS) is a surgical procedure for treating motor-related neurological disorders. DBS clinical efficacy hinges on precise surgical planning and accurate electrode placement, which in turn call upon several image processing and visualization tasks, such as image registration, image segmentation, image fusion, and 3D visualization. These tasks are often performed by a heterogeneous set of software tools, which adopt differing formats and geometrical conventions and require patient-specific parameterization or interactive tuning. To overcome these issues, we introduce in this article PyDBS, a fully integrated and automated image processing workflow for DBS surgery. METHODS: PyDBS consists of three image processing pipelines and three visualization modules assisting clinicians through the entire DBS surgical workflow, from the preoperative planning of electrode trajectories to the postoperative assessment of electrode placement. The system's robustness, speed, and accuracy were assessed by means of a retrospective validation, based on 92 clinical cases. RESULTS: The complete PyDBS workflow achieved satisfactory results in 92 % of tested cases, with a median processing time of 28 min per patient. CONCLUSION: The results obtained are compatible with the adoption of PyDBS in clinical practice.

Automatic computation of electrode trajectories for Deep Brain Stimulation: a hybrid symbolic and numerical approach.

PURPOSE: The optimal electrode trajectory is needed to assist surgeons in planning Deep Brain Stimulation (DBS). A method for image-based trajectory planning was developed and tested. METHODS: Rules governing the DBS surgical procedure were defined with geometric constraints. A formal geometric solver using multimodal brain images and a template built from 15 brain MRI scans were used to identify a space of possible solutions and select the optimal one. For validation, a retrospective study of 30 DBS electrode implantations from 18 patients was performed. A trajectory was computed in each case and compared with the trajectories of the electrodes that were actually implanted. RESULTS: Computed trajectories had an average difference of 6.45° compared with reference trajectories and achieved a better overall score based on satisfaction of geometric constraints. Trajectories were computed in 2 min for each case. CONCLUSION: A rule-based solver using pre-operative MR brain images can automatically compute relevant and accurate patient-specific DBS electrode trajectories.