Introduction
The birth of stereotaxic surgery dates back to the early 20th century when Sir Victor
Horsley and Robert Clarke used an apparatus and cartesian coordinates to probe a subcortical
structure of an animal subject.[1] Later, with the serendipitous discovery of some destructive lobar procedures' role,
stereotaxy was extrapolated to human patients in 1947.[2] Since then, the stereotactic radiosurgery field has flourished with Dr. Lars Leksell's
game-changing contributions. It was rightly considered an avenue for functional preservation
and less invasive surgery. In its initial days, it was used for functional disorders;
however, later, its efficacy was established in other brain pathologies like vestibular
schwannomas, pituitary tumors, craniopharyngiomas, cavernous sinus hemangiomas, metastasis,
and other vascular malformations. Their application was also extended to conditions
like trigeminal neuralgia, movement disorders, mesial temporal lobe epilepsy, and
certain psychiatric illnesses like obsessive compulsive disorders. Over time, significant
technological advancements have been witnessed in Gamma Knife treatment in the form
of computerized planning and image guidance, making it one of the most sought-after
treatments.
The delivery of gamma radiation using the Gamma Knife involves a meticulous sequence
of events to ensure precision and safety. It starts with patient selection using suitable
imaging methods and identifying the right candidate for this form of management. A
stereotactic frame is placed on the day of the procedure, following which the patient
is subjected to appropriate imaging. The imaging data are uploaded into the Gamma
Knife console that fuses the diagnostic images with the frame-based images, allowing
the precise localization of the target in three-dimensional space. Several multiparametric
thin slice MR images (i.e., T1W, T1W + C, and T2W) characterize the target lesion's
anatomical details. Manual contouring of the target lesion is performed, and the structures
at risk are defined. The dose matrices are placed, and treatment fields are defined
and revised. The treatment parameters are verified, and the plan is exported to the
Gamma Knife table. The patient is docked to the treatment table, and the highly focused
gamma radiation from 6°Co sources is delivered to the target. Most of the described steps in the radiation
delivery have subjective involvement and thus have a margin of error. Artificial intelligence
(AI) can reshape Gamma Knife radiosurgery (GKRS) by improving imaging, treatment planning,
and posttreatment assessment ([Table 1]).
Table 1
Studies on the application of Artificial Intelligence in GKRS
|
Study
|
Study focus
|
Findings and conclusion
|
|
Buzea et al[6]
|
To investigate three deep learning models, namely, the CNN model, the TL model, and
the FT model
|
The CNN model, trained from scratch, provided the most accurate predictions of SRS
responses for unlearned BM images
|
|
Ranjbarzadeh et al[7]
|
To obtain a flexible and effective brain tumor segmentation system
|
Several machines and deep learning models, namely, C-CNN model, can accurately obtain
a segmentation result
|
|
Tangsrivimol et al[8]
|
Review effectiveness of AI in GKRS by reviewing the published literature
|
ML techniques have proven effective in tumor identification, surgical outcome prediction,
seizure outcome prediction, aneurysm prediction, and more, highlighting its broad
impact and potential in improving patient management and outcomes in neurosurgery
|
|
Shapey et al[10]
|
To develop a novel artificial intelligence (AI) framework to be embedded in the clinical
routine for automatic delineation and volumetry of VS
|
A robust AI framework for automatically delineating and calculating VS tumor volume
and have achieved excellent results, equivalent to those achieved by an independent
human annotator
|
|
Klinge et al[11]
|
To investigate the feasibility of employing inverse planning methods to generate treatment
plans exhibiting desirable BED characteristics using the per isocenter beam-on times
and delivery sequence
|
They demonstrated the feasibility of using an inverse planning approach within a reasonable
time frame to ensure BED-based objectives are achieved across varying treatment times
and highlight the prospect of further improvements in treatment plan quality
|
|
Khouy et al[13]
|
They propose a new approach called GA-U Net that employs genetic algorithms to automatically
design a U-shaped convolution neural network
|
GA-U Net, a more viable option for deployment in resource-limited environments or
real-world implementations that demand more efficient and faster inference times
|
|
Feng et al[14]
|
To develop a deep learning model using a 3D U-Net with adaptations in the training
and testing strategies, network structures, and model parameters for brain tumor segmentation
|
A high prediction accuracy in both low-grade glioma and glioblastoma patients was
seen using a deep learning model
|
|
Ahn et al[16]
|
A deep learning method for dose prediction was developed and was demonstrated to accurately
predict patient-specific doses for left-sided breast cancer. Using the deep learning
framework, the efficiency and accuracy of the dose prediction was seen
|
|
Goyal et al[18]
|
Artificial neural network (ANN) model could predict the outcomes in trigeminal neuralgia
with 90% accuracy
|
|
Ertiaei et al[20]
|
ANN can predict postoperative outcomes in patients who underwent GKRS with a high
level of accuracy
|
Abbreviations: BED, biologically effective dose; BM, brain metastases; CNN, convolutional
neural network; C-CNN, Cascade Convolutional Neural Network; FT, fine-tuning; GA,
genetic algorithm; GKRS, Gamma Knife radiosurgery; ML, machine learning; SRS, stereotactic
radiosurgery; TL, transfer learning; VS, vestibular schwannoma.
Enhanced Imaging and Diagnosis
Imaging is fundamental to GKRS, serving as the basis for diagnosis, treatment planning,
and execution. The success of the procedure depends largely on the precision of imaging
techniques like magnetic resonance imaging (MRI), computed tomography (CT), and positron
emission tomography (PET). Machine learning is a significant subset of AI, which refers
to a system's ability to gain statistical insights by deciphering patterns from training
data and applying learned rules to predict a specific task. Apart from the derivation
of rules for optimal behavior, it also adapts to evolving circumstances. Deep machine
learning, commonly known as deep learning (DL), is a subset of machine learning that
advances this process by allowing the computer to analyze the extracted patterns and
create complex visual representations.[3] This is achieved through a series of simple mathematical operations organized into
layers of progressively more abstract feature extraction mechanisms using various
neural networks. The amount of information that can be interpreted using DL methods
can be quite enormous and accurate, thus improving diagnostic accuracy and even predicting
outcomes based on preprocedure imaging.[4]
Convolutional neural networks (CNNs) are the most common type of DL model used, and
they have been shown to remove noise from MR images while preserving resolution, leading
to better visualization of tumors and abnormalities.[5] This improved clarity is crucial for accurate targeting during GKRS. To make the
adjuvant therapy more meaningful, lesion segmentation is used to derive crucial information
about the tumor. A region-based CNN is a family of methods that can be utilized for
object detection and localization in the images. It can divide an image into multiple
regions and process them individually to search for the desired pathology. This method
has shown its potential application for the prediction of brain metastasis.[6] Traditional methods are labor intensive and time-consuming. Several machines and
DL models, namely, multi-atlas registration algorithms and DL methods, can be trained
to accurately distinguish between tumor and healthy brain tissue, along with providing
clinicians with precise data on the treatment area's location, size, and shape, thus
defining the target accurately.[7] This supplements treatment planning and can lead to better patient outcomes.[8] AI also supports multimodal image fusion, which combines data from various imaging
techniques (such as MRI and PET) to offer a more comprehensive view of the target
area. AI algorithms can automate this fusion process, ensuring accurate image alignment
and increasing the radiosurgery's precision.[9] By delivering a detailed understanding of the biological characteristics of the
target tissue, AI-powered image fusion helps optimize radiation dose selection and
treatment strategies. Shapey et al have successfully demonstrated a robust AI framework
that has accurately annotated vestibular schwannomas in their cohort and has promising
potential for its application in other tumors.[10]
Optimized Treatment Planning
In GKRS, we intend to deliver focused radiation on the target lesion and prevent scattering
to surrounding normal tissue, thus minimizing radiation-related side effects on normal
structures. The delivery of radiation dose is influenced by the geometry of the radiation
unit, the patient positioning system, and the collimator selection.[11]
Traditionally, this process is complex and iterative, relying on the expertise of
medical physicists and neurosurgeons. AI can significantly change treatment planning
by enhancing predictive modeling, automating contouring, and improving dose calculation.
AI algorithms, particularly reinforcement learning and DL, can create predictive models
that simulate treatment scenarios and forecast patient outcomes based on various parameters.
By analyzing vast amounts of historical data, these models can identify patterns and
correlations that may be obscure to human planners. Using deep neural network algorithms,
the dose distribution can be derived depending on the patient's anatomy and the beam
angles, and doses are adjusted accordingly.[12] Currently, in Gamma Knife, a semiautomated in-plane segmentation method is used,
which requires manually segmenting each axial slice, thus making the entire process
time-consuming. An automated segmentation tool using DL models can delineate the target
tumor tissue accurately and more quickly.[10] This approach is relatively time-consuming and subjective and could be made more
efficient by introducing an automated segmentation tool.
AI-based segmentation algorithms, such as U-Net, which is an artificial neural network
(ANN) architecture primarily used for segmentation in computer vision, have demonstrated
high accuracy in automating the contouring process.[13]
[14] AI reduces interobserver variability and promotes uniformity in treatment planning
by ensuring consistent and precise delineation of the target area and adjacent structures.
AI can also optimize dose calculation, a key component of GKRS planning. Traditional
methods, such as Monte Carlo simulations, are accurate but require significant computational
resources and time.[15] DL models can accurately compute dosage requirements based on the datasets of previous
cases, thus making the treatment delivery process more efficient and significantly
shorter in duration.[16]
Real-Time Monitoring and Adaptive Treatment
The efficacy of Gamma Knife is, apart from imaging and treatment planning, significantly
affected by controlled and precise delivery of radiation to the target, which is affected
by factors like patient movement, anatomical changes, and physiological variations.
AI-integrated systems can monitor patients' anatomy and position during the procedure
and mitigate these challenges.
For example, DL algorithms can analyze real-time imaging data to detect any shifts
in patient positioning or changes in the target location. If deviations are identified,
the system can alert the medical team or automatically adjust the treatment parameters
to maintain precise radiation delivery. This reduces the risk of radiation exposure
to healthy tissue, enhancing the procedure's safety and effectiveness.
Additionally, AI enables adaptive treatment planning, allowing the treatment plan
to be dynamically adjusted based on real-time data. This is particularly useful when
the target tissue or surrounding anatomy changes during the procedure. AI algorithms
can rapidly process new data, recalculate optimal treatment parameters, and update
the plan accordingly. Adaptive planning ensures more personalized and precise radiosurgery,
ultimately improving patient outcomes.[17] This can be of great use during administration of Gamma Knife for functional disorders,
namely, trigeminal neuralgia or performing ventral intermediate (VIM) thalamotomy
for control of tremors where the target tissue is of small volume and a large dose
of radiation is to be administered. Thus, precision and adaptive treatment would play
a vital role.
Posttreatment Assessment and Outcome Prediction
Predicting long-term patient outcomes is one of the critical aspects of management.
AI can potentially predict the treatment responses and outcomes based on the preoperative
scans and the pathology, which otherwise would involve long-term clinical and radiological
follow-ups. For instance, DL models can be trained to identify residual tumor tissue
or early signs of recurrence with high sensitivity and specificity, allowing for timely
interventions and adjustments to the treatment plan if necessary. Goyal et al assessed
an ANN model trained on 16 variables to forecast postoperative outcomes after GKRS
in patients suffering from trigeminal neuralgia. An accuracy of 90.9% in predicting
treatment responses was achieved using this model.[18] The use of ANNs in medicine began in the late 1980s, particularly for diagnosis,
assessing disease severity, and outcome prediction. While ANNs were conceptualized
before modern computers, advancements in computational models led to their rapid adoption.[19] Ertiaei et al used ANNs to predict outcomes in patients with trigeminal neuralgia
who underwent stereotactic radiosurgery and to categorize the relative importance
of individual risk factors.[20]
[21]
Moreover, AI can support the creation of personalized treatment strategies by identifying
patient-specific factors that influence treatment outcomes. By analyzing data from
a vast number of cases, AI can reveal patterns and correlations that may not be apparent
in smaller datasets. This insight can inform the development of tailored treatment
protocols, optimizing the likelihood of a successful outcome for each patient. A Chinese
group used machine learning models to their advantage to satisfactorily predict post-Gamma
Knife edema in patients with meningioma.[22] They could thus counsel their patients better, tailor their treatment decisions,
and generate a customized follow-up plan.
Challenges
AI has been quite instrumental when it was used for pathologies like vestibular schwannoma,
trigeminal neuralgia, or metastasis, but it does have its share of limitations. The
algorithms used for the above conditions have been used in a small cohort of patients,
and their generalizability and reproducibility are yet to be established using prospective
studies. This warrants collaborative efforts from clinicians across the globe and
validates its efficacy and reliability in diverse populations.
Adopting and seamlessly integrating the practice of AI demands substantial adjustments
from clinicians. Clinicians must be trained to incorporate these tools into daily
use and comprehend outputs from AI software. Infrastructural amendments in the form
of provision of secure networks, processing power, and data storage have to be carried
out. Also, concerns about maintaining privacy crop up. Above all, the analysis, interpretation,
and precision offered by AI are based on the data fed to build the algorithms. Their
interpretation and analysis can considerably differ from that of an astute clinician.
Implementation of new pathologies might also take a long time before one can confidently
implement them for patient care.[22] There are several ethical issues that have to be addressed before AI technology
is widely incorporated into routine practice, which include patients' consent to utilize
the medical data, safety and transparency, and algorithmic biases. The inaccuracy
resulting from the algorithmic biases in treatment delivery can prove catastrophic
for the patient, and the extent of the harm it can result is difficult to gauge. In
addition, we cannot hold anyone accountable for faulty decision-making. This also
warrants prior intimation to the patients for whom AI technology is going to be used,
and one should stress on obtaining proper consent.[23]
Despite these challenges, the future of AI in GKRS is auspicious. Ongoing advancements
in AI algorithms, computational power, and data accessibility are poised to fuel continued
innovation. Furthermore, the convergence of AI with other emerging technologies—such
as robotics and augmented reality—can potentially elevate precision and personalization
in radiosurgical procedures to new heights. We also believe that introducing AI can
jeopardize the age-old clinician–patient connection, which can affect patient trust,
and we will have to devise ideas to circumvent these issues.
Conclusion
AI is set to revolutionize GKRS by advancing imaging capabilities, refining treatment
planning, enabling real-time monitoring, and enhancing posttreatment evaluations.
Through the integration of AI, clinicians can attain unprecedented levels of precision,
efficiency, and personalization in radiosurgical procedures, significantly improving
patient outcomes, as highlighted in [Table 1]. Although challenges persist, the ongoing development and adoption of AI-driven
innovations offer tremendous potential for the future of GKRS and the broader discipline
of neurosurgery.
