Page 141 - AIH-2-4

Artificial Intelligence in Health Artificial intelligence app for EVD navigation

It is important to acknowledge, however, that not structures and pathologies (e.g., tumors, vasculature,
all practicing neurosurgeons will adopt or require this cranial nerves). With Apple’s CoreML architecture,
technology in their workflow. Surgeons with extensive the models are efficiently accelerated, facilitating rapid
clinical experience, including hundreds of EVD inferences that drive the app.
placements, may find little practical benefit in an additional The use of AR in neurosurgical navigation offers a more
technological layer. Indeed, the tactile and anatomical intuitive and immersive experience for the surgeon while
intuition developed over decades cannot be easily replaced minimizing the fatigue often associated with virtual reality
or replicated by software. This application is therefore solutions. The iOS application utilizes AR to provide
3
not intended to supplant surgical judgment; rather, it is real-time visualization of surgical trajectories, anatomic
designed to augment procedural safety and education, boundaries, and feedback regarding the system’s accuracy.
particularly for trainees, early-career providers, and those These cues may help provide surgeons with an “X-ray
who infrequently perform EVD placements.
vision–” like understanding of patient-specific anatomy,
Trainees often face steep learning curves in ventricular thereby increasing confidence during procedures.
catheterization, a task further complicated by anatomical While the present study focused on the feasibility of
variability, patient movement, and the urgency of emergent using an iOS application for navigated EVD placements,
settings. The traditional apprenticeship model, while there are many potential future applications for this
time-tested, provides variable exposure and feedback. By technology. For example, the iOS application could be
delivering real-time visual guidance, trajectory alignment,
and error detection, AI-based navigation can shorten used for remote surgical guidance, allowing a surgeon in
the time needed to achieve competence, reduce patient one location to guide another surgeon through a procedure
risk, and improve trainee confidence. Recent research using shared visualization. In addition, it could serve as
a training and educational tool, providing a realistic and
supports this potential, showing that access to simulation
and navigational feedback correlates with faster skill immersive simulation environment for trainees to practice
acquisition and lower complication rates in neurosurgical neurosurgical navigation or to explore anatomy through an
training programs. 51 alternative medium. The next step will be to compare the
application’s navigational accuracy against the gold standard
Registration remains the key step for mapping a patient’s of head CT for EVD placement in cadaveric models.
physical anatomy to preprocedural imaging. Conventional
systems often rely on as few as 10 points to generate a There are several limitations to this study that should be
rigid transformation between real-world and radiographic acknowledged. First, this is a proof-of-concept feasibility
coordinates. In contrast, our proprietary system leverages study. While our model successfully demonstrated that an
6
the full 30,000 points provided by Apple’s TrueDepth camera, iOS application can track an EVD in real time, its effect on
corrects for device-specific intrinsic properties, and performs EVD placement accuracy and clinical outcomes remains
transformations unique to each video frame. This allows unknown. Given the feasibility design, only a limited set of
anatomy to move relative to the camera while maintaining evaluation metrics and parameters were analyzed, which
continuous registration. These operations are accelerated by will be expanded in future studies. The next step will involve
the on-device GPU, minimizing computational burden on cadaveric testing to rigorously evaluate the accuracy and
52
the mobile device. Further study is warranted to quantify the safety of the iOS application in a controlled setting. In
impact on iOS battery performance. addition, the small sample size limits the generalizability
of findings, underscoring the need for larger studies.
Overall, the use of transfer learning allowed us to leverage Furthermore, the application’s overall clinical utility in
pre-trained models developed on large datasets to train navigating EVD placement cannot be determined until it
our models on relatively small datasets, resulting in high is directly compared with the gold standard of head CT,
accuracy and robustness. For the head CT segmentation which will be the subject of subsequent research.
model, we leveraged the well-studied open-source QURE.
AI CQ500 dataset, consisting of 491 head CTs obtained Nonetheless, the present study demonstrates the
on multiple scanners, which improved the generalizability potential of iOS devices to improve neurosurgical navigation,
of our model compared to training on institutional CTs particularly for trainees and inexperienced providers, and
alone. The data collection and annotation processes for establishes a foundation for future research in this area.
45
both models were time-consuming and required expert 5. Conclusion
knowledge. These models further provide opportunities for
further iteration and the incorporation of richer datasets, The goal of this investigation was to integrate existing
such as MRI, or for segmentation of additional anatomic technologies in registration and object tracking into a

Volume 2 Issue 4 (2025) 135 doi: 10.36922/aih.8195