Utility of Artificial Intelligence in Stereotactic Radiosurgery for Vestibular Schwannomas: A Systematic Review

Abstract

Vestibular schwannomas (VSs) are benign neoplasms commonly located in the cerebellopontine angle and are increasingly managed with stereotactic radiosurgery (SRS), particularly Gamma Knife radiosurgery (GKRS). The integration of artificial intelligence (AI), encompassing machine learning (ML) and deep learning (DL) algorithms, into GKRS has emerged as a promising strategy to enhance diagnostic accuracy, automate treatment planning, and predict treatment response. This systematic review evaluates the current applications and clinical utility of AI in the stereotactic radiosurgical management of VSs. A systematic search was conducted on July 31, 2024, across Medline (PubMed), Embase, Scopus, and the Cochrane Library, in accordance with PRISMA guidelines. Studies were selected if they investigated the use of AI at any stage of stereotactic treatment or follow-up of VSs. Articles were excluded if they focused solely on microsurgical interventions or were review articles. Eligibility was independently assessed by two reviewers, with discrepancies resolved by a third observer. A total of 22 original studies were included in the final qualitative synthesis. AI applications were categorized into three domains: (1) pre-treatment tumor characterization and segmentation, (2) radiosurgical treatment planning, and (3) post-treatment response prediction. Multiple studies demonstrated the efficacy of convolutional neural networks (CNNs) and federated learning for automated and accurate segmentation of VSs, often achieving performance metrics comparable to expert manual annotations. In treatment planning, AI-driven models enabled improved target delineation, dosimetric optimization, and reduced inter-planner variability. In the post-treatment phase, radiomic-based AI models accurately predicted pseudoprogression and long-term tumor response, while automated volumetric assessment tools reliably tracked tumor changes over time. Collectively, these AI applications showed potential to streamline clinical workflows, enhance precision, and support individualized decision-making. AI has shown significant promise in enhancing various aspects of stereotactic radiosurgical care for VSs, from diagnosis and planning to longitudinal monitoring. While current findings are encouraging, challenges such as data standardization, model generalizability, and integration into clinical practice remain. Further prospective multicenter studies and regulatory oversight are warranted to validate AI tools and facilitate their widespread clinical adoption. With continued refinement, AI is likely to augment the capabilities of radiosurgeons and improve outcomes for patients with VS.

Keywords

artificial intelligence - Gamma Knife - machine learning - radiomics - radiosurgery - vestibular schwannoma

Introduction

Vestibular schwannomas (VSs) are the most common benign tumors found in the cerebellopontine angle region.[1] [2] Most carry a good prognosis and can be treated either surgically or via stereotactic radiosurgery, with the latter gaining popularity nowadays. Stereotactic techniques include most commonly Gamma Knife radiosurgery (GKRS) and Cyber Knife therapy, and are applicable mostly to tumors less than 3 cm in size.[3] [4] [5] They can be classified either as primary therapy when given upfront or secondary therapy when given for residual lesions after surgery. Artificial intelligence (AI) describes computer technologies that are modeled after human intelligence processes. The ability of a system to gain statistical knowledge through pattern recognition in training data and rule-based rule-learning to forecast a given predetermined task based on these patterns is known as machine learning. This process is furthered by deep machine learning, or simply deep learning (DL), techniques, which allow the computer to create its own intricate and sophisticated visual representations using basic mathematical operations in addition to statistical reasoning on extracted patterns. In medical image analysis, convolutional neural networks (CNNs), also called networks by machine learning practitioners, are the most widely used DL models. They have attained state-of-the-art performance for numerous segmentation tasks. Such AI models are usually trained using a set of annotated training images under supervision. From individualized treatment planning to diagnostics, the introduction of AI in medicine has completely changed several facets of the healthcare industry. AI has the potential to improve treatment results, accuracy, and efficiency when used in conjunction with Gamma Knife therapy for VS patients. This review will provide a concise background and foray into ongoing, rapidly progressing AI research by compiling the current literature in a systematic fashion.

Materials and Methods

Search Strategy

A systematic search was performed in accordance with the PRISMA guidelines to identify and compile peer-reviewed articles highlighting the utility of AI in stereotactic radiosurgical treatment of patients with VS. Various machine algorithms and databases used in the studies were identified, and the stage of treatment at which the algorithm was used to aid in stereotactic treatment of VS was noted. The following search strategy was used:

Search Phrase

((“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” OR “Neural Network” OR “Dosimetric Analysis” OR “Radiation Planning” OR “Radiation Therapy Planning” OR “Radiosurgical Planning” OR “Radiomics”) AND (“Gamma Knife” OR “Stereotactic Radiosurgery” OR “Cyberknife” OR “Radiosurgery” OR “Radiotherapy”) AND (“Vestibular Schwannoma” OR “Acoustic Neuroma”))

All studies that displayed any of these terms in the title or abstract were collected for further review.

Database

Medline (PubMed), Embase, Scopus, and Cochrane Library (Wiley).

Timeline

The systematic search was performed on July 31, 2024, using the above databases.

Study Eligibility

Studies were considered eligible if they investigated the role of any AI or machine learning algorithm during any stage of the stereotactic radiosurgical treatment of VSs or in the posttreatment follow-up. English language original articles were included without geographic or temporal preference. Prior reviews were excluded. Study eligibility was assessed by two independent observers (A.G.K. and K.P.), and disagreements were resolved with the opinion of a third independent observer (S.V.).

Results

Literature search yielded 31 studies; they were screened through the title/abstract of study/availability of full text; and 23 relevant studies were assessed for eligibility. Eight studies were excluded as they mainly depicted the use of AI for microsurgical treatment of VSs and not for stereotactic radiosurgery. One additional study was identified from the references of studies found. Two reports were excluded for being previous reviews. A total of 22 studies were included for qualitative summarization ([Fig. 1]). Studies were stratified into three categories depending on the stage of stereotactic management of VSs which highlighted the objective utility of AI. [Fig. 1] shows the PRISMA model detailing our systematic review process.

Fig. 1 PRISMA diagram summarizing search strategy and results.[5] Search phrase: ((“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” OR “Neural Network” OR “Dosimetric Analysis” OR “Radiation Planning” OR “Radiation Therapy Planning” OR “Radiosurgical Planning” OR “Radiomics”) AND (“Gamma Knife” OR “Stereotactic Radiosurgery” OR “Cyberknife” OR “Radiosurgery” OR “Radiotherapy”) AND (“Vestibular Schwannoma” OR “Acoustic Neuroma”)). Database: Medline (PubMed), Embase, Scopus, and Cochrane Library (Wiley), Timeline: Till July 31, 2024.

Pre-Radio Surgical Characterization and Segmentation

Automatic segmentation of VS will help reduce the clinical workload and speed up the clinical decision-making and help in patient management. AI and machine learning models can be used right from the prediction of a lesion as VS from a radiological dataset, its radiological grading, and even the prediction of post–gamma knife response of the tumor. Lee et al,[6] in their study, proposed a DL model to effectively segment VSs, brain metastases, and meningiomas for patients undergoing gamma knife radiosurgery, which can be utilized for gamma knife planning. They used a CNN training model and established the superiority of training two parametric MR images for segmenting brain metastases and VSs. The main advantage was better delineation of regions with inhomogeneous signal intensities. A higher degree of identification than that of meningioma and metastases was facilitated by the tumor's consistent placement at the cerebellopontine angle and the heterogeneous contrast enhancement of VS. They even successfully implemented this lesion delineation framework into a federated learning (FL) framework, which solves decentralization and privacy concerns.[7] Windisch et al[8] similarly developed a model using CNNs to differentiate and recognize VSs, glioblastoma, or no tumor. Three datasets were utilized: healthy MRI slices from IXI dataset, VS images from the European Cyberknife Center in Munich, and glioblastoma slices from the Cancer Genome Atlas dataset. They used Grad-CAM, which achieved a categorical accuracy of 0.93 that further increased to 0.97 on implementing a Bayesian neural network. Shapey et al[9] were among the first to develop a robust AI-based framework for automatic segmentation of VSs, delineating and calculating the volume of tumor, aiding in treatment planning. Contrast-enhanced (CE) T1-weighted and high-resolution T2-weighted images were collected from all eligible patients who underwent stereotactic radiosurgery and utilized to create the model. They developed an AI model based on 2.5D CNN and compared it with manual segmentation to assess the accuracy. To evaluate the accuracy of the model, they quantitatively examined the dice score, average symmetric surface distance (ASSD), and relative volume error (RVE) of the segmentation results obtained from the automatic process as compared with the manual segmentations. The automated model was shown to produce results equivalent to the manual segmentation method with contrast-enhanced T1 images and CET1/T2 images. A 2.5D CNN benefits from enhanced in-plane feature extraction but exhibits coarse through-plane feature extraction. It has interslice abilities absent in 2D networks but requires less memory than 3D networks and is less complex, providing a midway path between 2D and 3D networks. McGrath et al,[10] in their study, compared the manual segmentation to semiautomated method for volumetric analysis of VSs. They found that semiautomated method was faster (167 vs. 479 seconds, p < 0.001), with increased accuracy and physically less intensive. Some limitations that could be pointed out were increased mental fatigue, algorithm unpredictability, and errors; however, they suggested semiautomated segmentation was comparable to manual and can be included in clinical practice. Semiautomated method offers the added advantage as compared with fully automated segmentation for being more transparent and accurate, rendering fully automated methods limited to academic interests. Shapey et al,[11] in their next study, provided an open-label, fully annotated dataset of 484 MRI images of patients undergoing gamma knife treatment at a single institution. Data also included contours and doses used for treatment planning. They used MONAI, an open-source framework utilized in healthcare imaging. Such datasets could be used to validate future segmentation models for VS and develop complex algorithms. Wang et al[12] similarly developed a 3D CNN model for fully automated segmentation of VSs based on T1-weighted MRI. It achieved good performance, validated on a good-sized dataset from two institutions. One of the datasets they included was that released by Shapey et al.[11] Milchenko et al,[13] in their study, analyzed the radiomic and clinical parameters predictive of AI segmentation for VSs. DeepMedic CNN architecture was utilized for the segmentation. They found that maximum and minimum signal intensity were the most predictive of tumor growth across all segmentation methods. AI can also be utilized for directly classifying the VSs which can have important clinical implications, can predict treatment planning, and reduce the decision time. Kujawa et al[14] developed the first AI framework model using CNN to classify VSs according to Koos grading. CNN was applied in two stages: a 3D U-Net CNN in the first stage and in the second stage, one branch was made to train either the DenseNet model or the Random Forest model, or an ensemble if both were used. Inter-rater reliability kappa was 0.68, and intra-rater reliability kappa of annotators 1 and 2 were 0.95 and 0.82, respectively. Their results established the accuracy of the model comparable to that of neurosurgeons and will aid patient management. Preradiosurgical radiomics can also be utilized to develop predictive models for postradiation response of VSs. Such models can be used to preemptively predict the radiological response of the tumor posttreatment: reduction, pseudoprogression, or stable size. Further course of treatment line expected could be decided beforehand. Yang et al[15] developed a machine learning model to predict pseudoprogression and long-term outcome in patients of VSs treated with Gamma knife. A total of 1,763 radiologic features were extracted from MRI sequences before GKRS to develop the model. Five radiomic parameters were considered based on T2-weighted and heterogenous contrast enhancement. Long-term outcome prediction achieved accuracy of 88.4% and pseudoprogression of 85%. Sümer et al,[16] in their study, utilized machine learning to discern the tumor shape, feature, and various morphological indices that can predict the gamma knife dose plan. They used tumor control and serviceable hearing preservation at 2 years as outcomes. Their plan quality was measured using indices like the selectivity index (SI), gradient index (GI), Paddick's conformity index (PCI), and efficiency index (EI). All tumor shape irregularity indices correlated significantly with the aforementioned indices. Volume index of sphericity was the most important predictor of the indices and had 89.36% accuracy for predicting PCI. However, treatment outcomes were not affected by tumor shape parameters. Similarly, Bossi Zanetti et al[17] developed a predictive network model for treatment response of VSs to Cyberknife therapy based on pretreatment MRI radiomics. They used four machine learning algorithms: random forest (RF), neural network (NNet), support vector machine (SVM) with radial kernel, and eXtreme Gradient Boosting (XGBoost) for testing the treatment response prediction. The neural network was the best algorithm for response at 24 (73% ± 18% balanced accuracy) and 36 months (65% ± 12% balanced accuracy). It showed more accuracy on how to separate patients with larger tumor volumes from the rest. Utilization of such models and machine learning algorithms can prevent unnecessary treatment and long follow-up of patients. Most of the studies have demonstrated the utility of automatic segmentation on the pr planning MRI datasets. Such datasets are usually standardized; diagnostic clinical datasets pose a bigger challenge. Kujawa et al[18] utilized DL framework to automate the segmentation of VSs on a routine clinical MRI dataset. They publicly released a multicenter routine clinical MRI dataset of 160 patients who had a single sporadic VS. A nnU-Net convolution network was used for training the models. Their model's dice similarity coefficients were similar to those achieved by trained radiologists when interobserver variation was noted. Three public VS datasets, one partially public dataset, and an MC-RC (multicenter routine clinical) hold-out testing set were used to assess the performance of the model. After being trained on the MC-RC dataset, the VS DL segmentation models' robustness and generalizability significantly improved. However, there was poor generalization among models trained on the Gamma Knife dataset. This signifies the importance of data variability in developing robust segmentation models.

Utilization in Treatment Planning

A step further in the utility of AI for VSs is the stereotactic radiosurgery planning. It can effectively segment the tumor, providing better delineation of the target. These algorithms can be utilized in the day-to-day clinical practice, providing assistance to the treatment planners. It can reduce workload significantly, especially in a high-volume center. The next added advantage is the reduction of inter-planner variability, maintaining the plan quality. The clinicians can spend more time on sophisticated cases with classical pathologies being planned with AI algorithms, albeit with manual validation if required. Lee et al,[19] in their study, developed a two-pathway model for better target delineation, which could aid in treatment planning. The two-pathway model fared better than the single pathway one in terms of better dice scores (0.90 ± 0.05 vs. 0.87 ± 0.07). Furthermore, the two-pathway model treated with biparametric (T2, T1 contrast) and triparametric (T1, T2, and T1 contrast) images had better dice scores than compared with single parametric treatment. Their model effectively segmented heterogeneous tumor target into solid and cystic components, which can aid in planning. Liu et al[20] developed one of the first automated inverse gamma knife treatment planning algorithms for VSs. They built a priority tuning policy network using deep CNNs. Multiple gamma knife plan metrics were used in their institution for plan evaluation, and the network was trained using a deep reinforcement learning (DRL) framework. Their network achieved scoring results comparable to expert manual treatment planners, indicating their potential utility in routine clinical management and treatment planning. The original plans with the identical starting priority set had average scores of 3.63 ± 1.34, 3.83 ± 0.86, and 4.20 ± 0.78. However, after manual priority adjustment by human expert planners, their scores increased to 5.28 ± 0.23, 4.97 ± 0.44, and 5.22 ± 0.26. With 5.42 ± 0.11, 5.10 ± 0.42, and 5.28 ± 0.20, respectively, their network produced competitive performance. Improvisation in dosimetric planning can also be achieved with AI utility. Safari et al[21] established this in their study by predicting and correcting the patient-specific distortions in the MRI images and the importance of doing so in the treatment planning. Their proposed method was fast, cheap, with easy implementation, and could be applied to diverse MRI datasets. Such techniques could be applied to routine clinical practice and will improve the standard of care in GKRS treatment and planning.

Predicting Treatment Response Based on Posttreatment Follow-up

Several studies have been done for response prediction based on tumor radiomics and morphological characteristics. Predictors can be determined regarding which tumors will show transient enlargement and which will respond without pseudoprogression. Usually, larger tumors have an initial pseudoprogression post-GKRS. This will enable clinicians to select tumors on an individual basis and decide the optimal treatment strategy between gamma knife therapy and surgical intervention. Langenhuizen et al,[22] in their study, utilized MRI tumor texture parameters for pseudoprogression prediction. Their study was prospective and utilized MRI images of patients taken at the time of treatment and at 6, 12, 24, 36 months thereafter. MRI features extracted were first-order statistics, Minkowski functionals (MFs), and three-dimensional Gray-level co-occurrence matrices (GLCMs). These were then studied in a machine learning algorithm for classifying pseudoprogression. Support vector machine (SVM) model was used. Due to the limited dataset, SVM proved helpful as it is efficient in binary classification problems and does not require a large amount of data. They determined that patient and treatment-related characteristics do not correlate with tumor enlargement. However, 4 GLCM features showed their prognostic value of pseudoprogression and the sensitivity and specificity increased with larger tumor sizes, more than 6 cm³. In their retrospective cohort study also, Langenhuizen et al[23] established GLCMs which extract features from MRI images as having the best prediction scores for tumor response. This further improved for tumors more than 5 cm³ in size. George-Jones et al,[24] also in their study, developed a model utilizing SVM to predict post-GKRS tumor enlargement. The model had a sensitivity of 92%, specificity of 65%, area under the curve of 0.75, and a positive likelihood ratio of 2.6 (95% CI: 1.4–5.0) for predicting post-SRS enlargement of >20%. It demonstrated an 87% sensitivity, 73% specificity, 0.76 AUC, and 3.2 (95% CI: 1.2–8.5) positive likelihood ratio in the bigger tumor subgroup and a sensitivity of 95%, specificity of 50%, AUC of 0.65, and positive likelihood ratio of 1.9 (95% CI: 0.8–4.3) in the smaller tumor subgroup. Huang et al[25] segmented the VSs into solid and cystic types based on retrospective MR imaging analysis of patients who received GKRS. First, the tumor was segmented into solid and cystic types using fuzzy C-means clustering. The response was classified as non-pseudo progression and pseudoprogression/fluctuation. They established that the risk of pseudoprogression was more for solid tumors compared with cystic tumors (55 vs. 31%, p < 0.001). Also, a mean lower signal intensity (SI) in T2/T1 contrast images before GKRS was associated with a higher risk of pseudoprogression both for solid and cystic tumors. Such radiological features can preemptively predict the tumor response after GKRS. AI can also be utilized for post-GKRS treatment volumetric analysis. This can help in quick tumor volume assessment in longitudinal follow-up of patients who received GKRS. Lee et al[26] developed an AI algorithm utilizing U-net CNN to automate the volumetric measurement of GKRS-treated VSs using a series of parametric MR images. A sample of 861 patients who had undergone GKRS was taken, comprising a total of 1,290 MR examinations (T1 contrast and T2-weighted sequences). The relative volume differences between AI-based measurements and clinical measurements performed by expert radiologists at each follow-up point were +1.74%, −0.31%, −0.44%, −0.19%, −0.01%, and +0.26% regardless of the state of the tumor (progressed, pseudoprogressed, or regressed). The discrepancies between the clinical measures made by knowledgeable radiologists and the outcomes of the suggested AI model fell below the range considered clinically acceptable, that is, less than 1% at all time points except the initial one. This model is applicable for assessments conducted at long-term intervals after a range of therapeutic procedures.

[Table 1] summarizes various studies demonstrating the utility of AI in stereotactic radiosurgery for VSs.

Table 1 Summary of various studies depicting the utility of AI in stereotactic radiosurgical treatment of vestibular schwannomas

Authors	Type of study	Number of images/patients analyzed	Type of AI tool/other technology/model used	Prominent findings	AI used for
Lee et al[6]	Retrospective	506 VS, 1069 meningioma, 574 brain metastases	Convolutional neural network	For VS and brain metastases, the model trained using two parametric MR images outperformed that using single parametric images, achieving higher median dice coefficients	Pre-op segmentation
Lee et al[7]	Retrospective	506 and 118 vestibular schwannoma patients aged 15–88 and 22–85 from two institutes, respectively	Federated learning (FL)	The suggested lesion demarcation was effectively included in an FL framework. Each participating institute's SRS data could be used with the FL models, and on non-SRS datasets, the FL showed a mean dice coefficient that was similar to that of CL	Pre-op segmentation, solving privacy concerns
Windisch et al[8]	Retrospective	1,023 healthy, 388 VS, 336 GBM	Convolutional neural network, Grad-CAM and Bayesian neural network	The model achieved a categorical accuracy of 0.93, creating a Bayesian neural network increased to 0.97	Pre-op segmentation
Shapey et al[9]	Retrospective	246 patients	Convolutional neural network	Automated approach outperformed an annotator statistically when ceT1 photos were used as the only input (p = 4e − 13) and when ceT1/hrT2 images were combined (p = 7e − 18)	Pre-op segmentation and volumetric calculation
McGrath et al[10]	Retrospective	4 patients	ITK-SNAP for manual and ImFusion Labels for semi-automation	When compared to manual segmentation, the semi-automated approach performed roughly equally and was considerably faster (167 vs. 479 s, p <0.001), less physically and mentally taxing, and had modest accuracy gains	Pre-op segmentation and volumetric calculation
Shapey et al[11]	Retrospective	484 MR images collected on 242 consecutive patients who underwent GKRS	Convolutional neural network	Provided the first fully annotated publicly available imaging dataset of VS, which can develop segmentation models	Pre-op segmentation
Wang et al[12]	Retrospective	737 patients	Convolutional neural network	The model performed well in volumetry and VS segmentation on a sizable dataset from two institutions	Pre-op segmentation and volumetry
Milchenko et al[13]	Retrospective	158 patients	Convolutional neural network	The characteristics that held up the best across segmentations were determined to be homogeneity, robust maximum intensity, and sphericity. Across all segmentation techniques and subject cohorts, maximum and lowest intensities were most indicative of tumor growth	Tumor growth prediction and segmentation
Kujawa et al[14]	Retrospective	308 patients	Convolutional neural network	The ensemble model's accuracy score, weighted macro-averaged F1 score, and weighted macro-averaged mean absolute error (MA-MAE) were comparable to those of two neurosurgeons and hence developed the first AI model to classify VS by the Koos scale	VS classification
Yang et al[15]	Retrospective	336 patients	Two-level binary classification model and support vector machine (SVM)	The accuracy of the long-term result prediction was 88.4%, based on five radiomic variables, and the prediction of transient pseudoprogression was accurate to 85.0%	Predict long-term response and pseudoprogression
Sumer et al[16]	Retrospective	234 patients	Recursive feature elimination and SVM	In vestibular schwannomas, TSI was found to be a significant factor influencing the quality of the GK dose plan, not the treatment outcomes. VioS was the best measure of shape irregularity across multiple metrics to predict the quality of the GK plan in vestibular schwannomas	Predict dose plan based on tumor morphology
Zanetti et al[17]	Retrospective	108 patients	Random forest, support vector machine, neural network, and extreme gradient boosting	At 24 and 36 mo, the neural network proved to be the most accurate forecast model for response	Predict treatment response
Kujawa et al[18]	Retrospective	160 patients	nnU-Net convolution network	Dice similarity coefficients (DSCs) were similar to those achieved by trained radiologists	VS segmentation on the routine dataset
Lee et al[19]	Retrospective	516 patients	Double pathway U-net model	When segmenting VS using anisotropic MR images, the suggested two-pathway U-Net model performed better than the single-pathway U-Net model	Target delineation for treatment planning
Liu et al[20]	Retrospective	5 training, 5 validation, 16 testing	Deep reinforcement learning	For vestibular schwannoma instances, their network produced GK plans that were comparable in quality to those created by human planners	Inverse treatment planning for VS
Safari et al[21]	Retrospective	Three publicly available datasets: the vestibular schwannoma-SEG (VS) dataset (242 patients), the slow event-related fMRI designs dataset (62 patients), and the MPI-Leipzig mind-brain-body (318 patients)	3D convolution network	Correcting patient-specific geometrical distortion in MRI images enhances the accuracy of dosimetric planning for vestibular schwannoma	Treatment planning
Langenhuizen et al[22]	Prospective	99 patients	SVM	MRI tumor texture can predict transient tumor enlargement, especially for large-sized VS	Pseudoprogression prediction
Langenhuizen et al[23]	Retrospective	85 patients	Supervised machine learning	Gray-level co-occurrence matrices obtained the best prediction scores for tumor response, especially for tumors larger than 5 cm3	Pseudoprogression prediction
Jones et al[24]	Retrospective	53 patients	SVM	For predicting post-SRS enlargement of >20%, the model showed a sensitivity of 92%, specificity of 65%, and positive likelihood ratio of 2.6	Pseudoprogression prediction
Huang et al[25]	Retrospective	323 patients	Fuzzy c-means	On T2W/T1WC images, tumors with greater tumor mean SI, solid component mean SI, and cystic component mean SI were more likely to have volume regression following GKRS	Response prediction
Huang et al[26]	Retrospective	330 patients	Fuzzy c-means	In contrast to cystic VS, pseudoprogression is more likely to develop in solid VS	Pseudoprogression prediction
Lee et al[27]	Retrospective	381 patients	Convolutional neural network	At each follow-up time point, the relative volume difference (RVD) between AI-based volumetric measurements and clinical measurements conducted by skilled radiologists was +1.74%, −0.31%, −0.44%, −0.19%, −0.01%, and +0.26%	Posttherapy volumetric measurement

Abbreviations: CL, centralized learning; GBM, glioblastoma multiforme; SI, signal Intensity; T1WC, T1-weighted contrast; T2W, T2-weighted imaging; TSI, tumor shape irregularity; ViOS, volumetric index of sphericity; VS, vestibular schwannoma.

Discussion

Gamma Knife therapy, a form of stereotactic radiosurgery, has emerged as a leading noninvasive treatment option for VSs, offering the advantage of being minimally invasive with minimal damage to surrounding tissues. Integrating AI into GKRS can open new avenues for the treatment of this benign yet disabling pathology. In numerous important areas, AI has the potential to completely transform gamma knife therapy. This is especially true when it comes to machine learning and DL algorithms. Some classical AI algorithms include CNN, recurrent neural networks (RNN), and generative adversarial networks (GAN); all three are mainly employed in imaging segmentation and quick analysis of MRI and CT imaging.

AI supports the development of incredibly individualized treatment regimens by accurately segmenting the tumor and surrounding vital components. Radiation dosage distribution can be optimized using machine learning models such as predictive modeling and reinforcement learning, ensuring maximal tumor targeting while preserving healthy tissue. This lessens adverse effects and improves tumor control. The next important role of AI algorithms can be in predictive analytics. AI can forecast healthcare outcomes by examining big patient treatment databases. Clinicians can make well-informed judgments by using this information to estimate tumor shrinkage, hearing preservation, and probable consequences. By utilizing predictive models, AI can suggest tailored treatment plans according to the unique features of each patient and the biology of the tumor, thereby improving the therapy's overall effectiveness. AI can enable real-time monitoring and adaptation of treatment. AI technologies are able to track the motions and physiological changes of patients in real time during the procedure. This flexibility enables prompt modifications to the treatment plan, ensuring maximum precision and security. Additionally, by continuously learning from ongoing treatments, it makes adaptive radiotherapy possible. This flexibility ensures that any differences in the anatomy of the patient or tumor responses are taken into consideration, thereby preserving the therapy's accuracy. Added to this AI simplifies several facets of Gamma Knife therapy, including selecting the most optimum plan, and image acquisition. Automation speeds up treatment center throughput while cutting down on the amount of time needed for treatment planning. AI-driven automated quality assurance systems minimize human error and improve treatment reliability by ensuring that all tools and processes meet the highest standards. A few recently emerging AI technologies include models like FL and explainable AI (XAI). These seek to preserve patient privacy and enable transparency, aiding in clinical decision-making. In the earlier literature review, several case studies have demonstrated the utility of AI in stereotactic radiosurgery for VS. These include studies right from enabling better tumor segmentation and target delineation, improved planning accuracy, prediction of treatment response beforehand based on imaging features, and even postprocedural volumetric calculations.

However, the inclusion of AI is faced with several challenges that need to be foreseen and overcome. Provision of highly annotated big datasets to be used in AI algorithms is often limited. Data consistency is also an issue that needs to be addressed. Privacy and security risks significantly increase when handling and processing large amounts of sensitive patient data. Ensuring regulatory body compliance is difficult but necessary. Gaining access to a variety of datasets frequently requires institutional collaboration, but sharing data while protecting privacy is a challenging problem. AI usually performs well on training data but are limited when applied to generalized unseen data. Therefore, they must be rigorously tested before deployment in clinical settings. Many AI models, especially DL algorithms, function as “black boxes” with limited transparency in their decision-making, which can make clinicians reluctant to trust and use them. Therefore, building AI models with transparent and comprehensible decision-making procedures is essential to winning over clinicians and guaranteeing patient safety. In addition, a robust interdisciplinary coordination is needed between clinicians, radiologists, data scientists, and engineers, which can be challenging to coordinate and manage. Incorporating AI networks into clinical workflow might be disruptive initially, and even their slightest inclusion is very time-consuming and requires robust clinical evidence to prove their benefits. Lastly, patient acceptance is the key that must be explained thoroughly regarding their treatment procedure being done with an AI-based algorithm. It is crucial to establish trust in the patients by being open and honest about the advantages and drawbacks of AI. Such limitations can be addressed by improving data quality and access, enhancing transparency, proper ethical regulation, interdepartmental coordination, and proper awareness and education of the patients.

Nevertheless, future prospects for AI in Gamma Knife therapy for VSs seem promising. Treatment precision, efficacy, and safety are expected to improve with continued research and technology developments. AI will serve as a corollary to industrialization by automating processes and reducing human error. However, it is unlikely to entirely replace Gamma Knife radiosurgeons; instead it will enhance the efficiency of GK centers by acting as a valuable assistant. Conducting multicenter trial studies and obtaining proper funding is hence essential to further garner clinical evidence to inculcate AI models in clinical practice.

Conclusion

The integration of AI in GKRS for VSs represents a significant advancement in medical technology. It can provide for better patient outcomes by streamlining procedures through the improvement of treatment planning, automation, real-time monitoring, and predictive analytics. Even though there are challenges, further study and technical developments should make it possible to overcome them and provide more individualized, effective, and precise care. The prospects of AI in GKRS appear promising, with its scope expanding to include additional lesion types such as meningiomas and arteriovenous malformations. Robust clinical trials and studies are, however, required in this regard. This review will offer a succinct overview and a glimpse into the fast-advancing field of AI research by methodically synthesizing the available literature.

Conflict of Interest

None declared.

^† Dr. K.P. and Dr. A.G.K. contributed equally to the writing of this manuscript.

• References

• 1 Carlson ML, Link MJ. Vestibular schwannomas. N Engl J Med 2021; 384 (14) 1335-1348
• 2 Babu R, Sharma R, Bagley JH, Hatef J, Friedman AH, Adamson C. Vestibular schwannomas in the modern era: epidemiology, treatment trends, and disparities in management. J Neurosurg 2013; 119 (01) 121-130
• 3 Goldbrunner R, Weller M, Regis J. et al. EANO guideline on the diagnosis and treatment of vestibular schwannoma. Neuro-oncol 2020; 22 (01) 31-45
• 4 Germano IM, Sheehan J, Parish J. et al. Congress of neurological surgeons systematic review and evidence-based guidelines on the role of radiosurgery and radiation therapy in the management of patients with vestibular schwannomas. In: Clinical Neurosurgery. Oxford University Press; 2018: E49-E51
• 5 Haddaway NR, Page MJ, Pritchard CC, McGuinness LA. PRISMA2020: An R package and Shiny app for producing PRISMA 2020-compliant flow diagrams, with interactivity for optimised digital transparency and Open Synthesis. Campbell Syst Rev 2022; 18 (02) e1230
• 6 Lee WK, Yang HC, Lee CC. et al. Lesion delineation framework for vestibular schwannoma, meningioma and brain metastasis for gamma knife radiosurgery using stereotactic magnetic resonance images. Comput Methods Programs Biomed 2023; 229: 107311
• 7 Lee WK, Hong JS, Lin YH. et al. Federated learning: a cross-institutional feasibility study of deep learning based intracranial tumor delineation framework for stereotactic radiosurgery. J Magn Reson Imaging 2024; 59 (06) 1967-1975
• 8 Windisch P, Weber P, Fürweger C. et al. Implementation of model explainability for a basic brain tumor detection using convolutional neural networks on MRI slices. Neuroradiology 2020; 62 (11) 1515-1518
• 9 Shapey J, Wang G, Dorent R. et al. An artificial intelligence framework for automatic segmentation and volumetry of vestibular schwannomas from contrast-enhanced T1-weighted and high-resolution T2-weighted MRI. J Neurosurg 2019; 134 (01) 171-179
• 10 McGrath H, Li P, Dorent R. et al. Manual segmentation versus semi-automated segmentation for quantifying vestibular schwannoma volume on MRI. Int J Comput Assist Radiol Surg 2020; 15 (09) 1445-1455
• 11 Shapey J, Kujawa A, Dorent R. et al. Segmentation of vestibular schwannoma from MRI, an open annotated dataset and baseline algorithm. Sci Data 2021; 8 (01) 286
• 12 Wang H, Qu T, Bernstein K, Barbee D, Kondziolka D. Automatic segmentation of vestibular schwannomas from T1-weighted MRI with a deep neural network. Radiat Oncol 2023; 18 (01) 78
• 13 Milchenko M, Cross K, Smith H. et al. AI segmentation of vestibular schwannomas with radiomic analysis and clinical correlates. . medRXiv 2023;
• 14 Kujawa A, Dorent R, Connor S. et al. Automated Koos classification of vestibular schwannoma. Front Radiol 2022; 2: 837191
• 15 Yang HC, Wu CC, Lee CC. et al. Prediction of pseudoprogression and long-term outcome of vestibular schwannoma after Gamma Knife radiosurgery based on preradiosurgical MR radiomics. Radiother Oncol 2021; 155: 123-130
• 16 Sümer E, Tek E, Türe OA. et al. The effect of tumor shape irregularity on Gamma Knife treatment plan quality and treatment outcome: an analysis of 234 vestibular schwannomas. Sci Rep 2022; 12 (01) 21809
• 17 Bossi Zanetti I, De Martin E, Pascuzzo R. et al. Development of predictive models for the response of vestibular schwannoma treated with Cyberknife^®: a feasibility study based on radiomics and machine learning. J Pers Med 2023; 13 (05) 808
• 18 Kujawa A, Dorent R, Connor S. et al. Deep learning for automatic segmentation of vestibular schwannoma: a retrospective study from multi-center routine MRI. Front Comput Neurosci 2024; 18: 1365727
• 19 Lee WK, Wu CC, Lee CC. et al. Combining analysis of multi-parametric MR images into a convolutional neural network: precise target delineation for vestibular schwannoma treatment planning. Artif Intell Med 2020; 107: 101911
• 20 Liu Y, Shen C, Wang T. et al. Automatic inverse treatment planning of Gamma Knife radiosurgery via deep reinforcement learning. Med Phys 2022; 49 (05) 2877-2889
• 21 Safari M, Fatemi A, Afkham Y, Archambault L. Patient-specific geometrical distortion corrections of MRI images improve dosimetric planning accuracy of vestibular schwannoma treated with gamma knife stereotactic radiosurgery. J Appl Clin Med Phys 2023; 24 (10) e14072
• 22 Langenhuizen PPJH, Sebregts SHP, Zinger S, Leenstra S, Verheul JB, de With PHN. Prediction of transient tumor enlargement using MRI tumor texture after radiosurgery on vestibular schwannoma. Med Phys 2020; 47 (04) 1692-1701
• 23 Langenhuizen PPJH, Zinger S, Leenstra S. et al. Radiomics-based prediction of long-term treatment response of vestibular schwannomas following stereotactic radiosurgery. Otol Neurotol 2020; 41 (10) e1321-e1327
• 24 George-Jones NA, Wang K, Wang J, Hunter JB. Prediction of vestibular schwannoma enlargement after radiosurgery using tumor shape and MRI texture features. Otol Neurotol 2021; 42 (03) e348-e354
• 25 Huang CY, Peng SJ, Yang HC. et al. Association between pseudoprogression of vestibular schwannoma after radiosurgery and radiological features of solid and cystic components. Neurosurgery 2023; 93 (06) 1383-1392
• 26 Lee CC, Lee WK, Wu CC. et al. Applying artificial intelligence to longitudinal imaging analysis of vestibular schwannoma following radiosurgery. Sci Rep 2021; 11 (01) 3106
• 27 Huang CY, Peng SJ, Wu HM. et al. Quantification of tumor response of cystic vestibular schwannoma to Gamma Knife radiosurgery by using artificial intelligence. J Neurosurg 2021; 136 (05) 1298-1306

Publication History

Article published online:
05 December 2025