U6 - VISION - Vessel Segmentation Maps Driving Automated Needle Insertion

Introduction: : Brain tumors present significant clinical challenges due to their locations and the intricate nature of brain anatomy. Intracranial metastases typically occur within the brain parenchyma, which is composed of functional tissue made up of neurons and glial cells. Glioblastomas, known for being the most aggressive, form in glial tissue and pose substantial treatment obstacles. Biopsies for diagnosis and medication delivery are essential for effective treatment, which requires visualizing and mapping the brain's vascular structure to minimize patient risk. While magnetic resonance angiography (MRA) is commonly used to visualize these vessels and help physicians plan trajectories that avoid veins and arteries, the complexity of the brain's vascular architecture still poses significant risks. Robotic needle insertion can enhance accuracy and reduce patient risk, particularly through steerable needles—novel medical devices that can curve when inserted into tissue, allowing navigation around critical anatomy. However, as their curvature has limitations and they are challenging for human operators to control in a 3D environment, their deployment is typically automated based on preoperatively determined motion plans. Creating and optimizing such plans with respect to patient safety represents a critical procedural step. This study introduces a novel approach using deep learning–based image analysis techniques to create comprehensive risk maps of the brain that quantify the likelihood of vessel presence. Our method leverages these risk maps to significantly improve the safety evaluation of motion plans for robotic-assisted needle insertion. This innovative strategy advances the safety of brain biopsy procedures and enhances targeted treatment options.

Materials and

Methods: : We use novel deep learning methods to create risk maps based on blood vessel segmentations. More specifically, we train a VesselBoost model, a Python-based deep learning model using a U-Net architecture, that creates highly accurate brain vessel segmentations from MRA scans. We use our trained model to generate detailed vascular maps by inputting MRA scans alongside their corresponding vessel labels. This rigorous training process allows the model to capture intricate vascular structures, enhancing the accuracy and reliability of the generated risk maps. Furthermore, we use Test Time Augmentation (TTA) strategies to refine the final segmentation results.
Deep learning models trained for segmentation tasks typically predict the probability of each voxel to be part of the structure to be segmented. A threshold is then applied to extract a binary segmentation. While we train our model for a standard segmentation task, at prediction time, we extract the pre-threshold probabilities to use as a risk map. This allows us to create optimized motion plans reducing the risk of crossing blood vessels. We use a Rapidly Exploring Random Tree (RRT) motion planner which evaluates each motion plan based on the risk map by summing up the probabilities for the presence of vessels for each voxel the path crosses. This approach enhances motion planning by prioritizing safety through a risk-aware evaluation, ultimately leading to more reliable and effective navigation in sensitive environments.

Results, Conclusions, and Discussions:: We used a set of 10 high resolution 7T MRA images and their corresponding vessel segmentation labels provided by the 2023 SMILE-UHURA Challenge to train our segmentation model, keeping two additional images separate to later be used for model testing purposes. To evaluate segmentation performance, we compared each voxel in the basic model prediction and TTA output images with the corresponding voxel in the ground truth segmentation labels. The trained model demonstrated high performance in segmenting vessels on the withheld ground truth images. The model’s predicted labels achieved an accuracy of 97%, precision of 98%, and recall of 99%, reflecting its ability to correctly identify vessel structures with minimal false negatives. The IoUscore was 97%, further confirming strong overlap between predicted and true vessel regions. Additionally, the F1 score reached 99%, highlighting the model’s overall robustness in balancing precision and recall.Applying additional TTA yielded similar results of 97% accuracy, 97% precision, 99% recall, 97% IoU, and 98% F1 score.
To evaluate the application to motion planning, we sampled 10 random tumor locations in the brain parenchyma of an MRA image and we ran the RRT motion planning algorithm to find motion plans to each location. In previous work, motion plans are often evaluated according to their length as a proxy metric for deployment risk.. Our simulations show that when comparing planning results using our new risk metric with those planned with the standard path length metric our new method reduces patient risk by 16.9%. When choosing the least risky plans according to our risk map, they only become 7.6% longer than the shortest plans on average.
These results demonstrate the feasibility of deep learning-based vessel segmentation for aiding brain biopsies and targeted treatment planning. Moving forward, our goal is to develop a versatile tool capable of processing any image type and resolution of different organs and to segment their vessels accurately. We aim to provide detailed vascular maps for physicians, which will enhance the safety of treatments and improve accessibility across various healthcare settings.

Acknowledgements and/or References (Optional):: [1] Chatterjee, Soumick, et al. "SMILE-UHURA Challenge--Small Vessel Segmentation at Mesoscopic Scale from Ultra-High Resolution 7T Magnetic Resonance Angiograms." arXiv preprint arXiv:2411.09593 (2024).
[2] Xu M, et al. "VesselBoost: A Python Toolbox for Small Blood Vessel Segmentation in Human Magnetic Resonance Angiography Data." Aperture Neuro (2024).