Enhancing Glioma Progression Prediction: Unsupervised Clustering of Whole-Brain Magnetic Resonance Spectroscopy Data for Surgical Planning

Abstract

Background:

The inevitable progression of high-grade gliomas has prompted a need for data-backed identification of compromised tissue prior to detection on traditional serial imaging. Whole-brain magnetic resonance spectroscopy (WB-MRS) can fill this role to classify glioma progression prior to contrast enhancement. Voxel-level data can differentiate areas of perilesional tissue under supervised machine learning (ML) and has been shown to predict the likelihood of tumor progression within six months. In this study, we aim to improve the spatial utility of WB-MRS ML through unsupervised ML clustering for utility in intraoperative integration with existing neuronavigation platforms.

Methods:

This analysis involved 16 adult patients that developed recurrence of high-grade glioma (HGG) on serial imaging, including 13 with glioblastoma (GBM) and three with anaplastic astrocytomas. Postoperative WB-MRS images were used as data inputs. We investigated two unsupervised clustering methods to optimize the identification of compromised perilesional tissue from a previously published supervised model. All voxels within the new region of interest (ROI) are reclassified as future progression regardless of their classification from the supervised model.

Results:

After hyperparameter tuning, Density-Based Spatial Clustering of Applications with Noise (DB-SCAN) shows an area under the curve (AUC) of 0.942, while K-Means Clustering shows an AUC of 0.874. DB-SCAN also shows a superior accuracy, specificity, sensitivity, and F1-score to the original supervised model alone.

Conclusion:

The incorporation of unsupervised clustering enhances the utility of WB-MRS for identifying compromised perilesional tissue and predicting high-grade glioma progression. Unsupervised clustering incorporates critical geospatial data and offers a more surgically relevant approach to visualizing tumor progression.

Importance of Study

Early detection of high-grade glioma recurrence remains a major limitation of current imaging modalities, restricting the ability to guide timely and precise interventions. While prior work has demonstrated that WB-MRS combined with supervised ML can identify metabolically abnormal tissue at risk for progression, these approaches lack spatial coherence and clinical usability. In this study, we introduce a hybrid pipeline integrating supervised voxel-wise predictions with unsupervised spatial clustering (DBSCAN and K-means) to refine recurrence mapping. This approach significantly improves classification performance and generates spatially contiguous, clinically interpretable regions of interest. Importantly, these outputs can be exported as DICOM overlays, enabling direct integration into neuronavigation systems. While still in pilot phase, this work advances WB-MRS from a predictive tool toward a clinically integrable platform, with potential applications in surgical planning, biopsy targeting, and longitudinal disease monitoring in patients with high-grade glioma.

Introduction

The persistent challenge of high-grade glioma (HGG) recurrence has driven recent research to explore nuanced approaches to the early detection of disease progression. Rising trends in machine learning (ML) across the field of both neuroradiology and neurooncology reveal a key role for radiomics, such as whole-brain magnetic resonance spectroscopy (WB-MRS), in identifying tissue at risk of tumor progression before it becomes radiographically evident on contrast-enhanced imaging^1–2. Such modeling could enhance expectant treatment planning and lead to improved morbidity and survival outcomes for this vulnerable patient population. Despite maximal safe surgical resection plus chemoradiation, outcomes for patients with glioblastoma (GBM) has remained stagnant over the last 20 years, with a median overall survival of 12–14 months³, and median progression free survival of 6–7 months⁴.

Through the quantification of the metabolites choline (Cho), creatine (Cr), n-acetyl-aspartate (NAA), glutamine-glutamate (Glx), and myo-inositol, WB-MRS enables locoregional detection of metabolic alterations associated with cellular division, metabolism, and the tumor microenvironment^5–11. These metabolic markers provide pathophysiologic insight that is otherwise not available in traditional radiomics analysis, allowing for the early detection of non-contrast enhanced yet at-risk tissue regions. Of note, evidence from our prior work⁵ and literature at large¹² suggests that the Cho/Cr ratio is the most influential marker for predicting progression, while the Cho/NAA^8,13 ratio has been used as an accepted marker of GBM.

Although WB-MRS has proven effective at identifying metabolic changes associated with glioma progression, translating voxel-level data to the intraoperative setting to improve surgical decision making remains a significant challenge. I our previous work forming the Miami Glioma Prediction Map (MGPM) (5) We demonstrated that WB-MRS combined with supervised ML could classify regions of HGGs that were at risk for progression within six months. Using this model, an area under the curve (AUC) with 5-fold cross-validation of 0.86±0.08 was reported, reaching 70.6% accuracy with the validation cohort.

Despite this, the MGPM assessed voxels solely on their metabolite quantifications, without incorporating their spatial location. Furthermore, while the model provides preoperative guidance via its publicly accessible web-based tools, the MGPM does not offer translational conversion to an intraoperative neuronavigation tool to visualize radiographically occult tumor.

In this study, we aim to address these limitations by exploring the use of unsupervised clustering methods, specifically Density-Based Spatial Clustering of Applications with Noise (DBSCAN)¹⁴ and K-Means¹⁵, to improve the spatial relevance and clinical utility of WB-MRS predictions. Both algorithms rely on spatial relationships to group data, allowing for the identification of meaningful patterns that may not be immediately apparent through analysis of individual voxels. Unsupervised clustering allows for the optimization of the supervised model, minimizing noisy false positive predictions distal to the core lesion and reclassifying false-negative voxels within areas consumed by compromised tissue. By incorporating spatial data, we aim to generate more precise and clinically relevant heat maps that can inform clinical decision making.

1.2. Methods

1.2.1. Patient Cohort

An overview of the methods workflow is visualized in Figure 1. A retrospective study evaluated WB-MRS data from 16 patients with stable disease who underwent WB-MRS scans status post resection of HGGs and whose disease progressed on surveillance imaging within eight months of the WB-MRS scan. The final cohort included 9 males and 7 females over the age of 18, with 13 glioblastoma (GBM) pathologies and 3 anaplastic astrocytomas included in the analysis. The mean time to recurrence after WB-MRS scan was 4.39 months. Full demographic data has been previously reported [5]. The review of patient images for this study was approved by the University of Miami Institutional Review Board (IRB #20160437). Consent for WB-MRS was obtained explicitly. Consent was waived for the review of medical records due to the lack of prospective patient contact.

Figure 1:

Flow chart outlining workflow. Patients with identified GBM undergo surgical resection and histologic confirmation of high-grade glioma. Postoperative WB-MRS is performed, and metabolites are used to train a Gradient Boosting supervised ML classifier. Unsupervised ML clustering is performed to identify high-risk voxels from the results of supervised predictions; the combined supervised and unsupervised maps are plotted over clinical MRI files and exported as DICOM files for neuronavigation software in the operating room. Created in https://BioRender.com.

1.2.2. Data Collection

MRI imaging data was collected from 2014 to 2019 under our institution’s clinical protocol⁵, and WB-MRS data was processed using The Metabolite Imaging and Data Analysis System (MIDAS)^7,16. Voxels of recurrence were outlined on surveillance imaging by board-certified neuroradiologists and hand-selected on correlated MRI sequences from the initial day of WB-MRS (months prior to recurrence) for use in supervised machine learning; for each scan evaluated, all slices were included in the review. These voxels are labeled as the area of future recurrence (AFR). The likelihood of recurrence for each voxel was calculated from a pre-trained Gradient Boosting model⁵, providing an output from 0% (no likelihood of recurrence) to 100% (highly likely to progress to tumor). This model was trained and tested with 14 patients; two were reserved for validation testing. The model requires all five inputs for predictive calculations.

1.2.3. Unsupervised Clustering

To incorporate spatial information and refine per-voxel predictions from the Gradient Boosting model, two unsupervised clustering algorithms, DBSCAN and K-Means, were evaluated. These clustering methods were implemented in Python (3.11) using sklearn.cluster packages. The primary objective was to cluster adjacent voxels exhibiting high-risk signatures as predicted by the supervised Gradient Boosting model within a surgical region of interest while reducing erroneous misclassifications distal to the tumor.

First, the voxel-wise outputs from the supervised Gradient Boosting model were plotted in three dimensional space; this map served as the input for the clustering algorithms. Next, both clustering models were implemented with varying hyperparameters in a grid-search fashion to to identify the settings of each model yielding the best performance (Table 1). For DBSCAN, we explored three hyperparameters: (1) the classification threshold, defined as the minimal predicted probability above which a voxel is designated high-risk; (2) minimum samples, referring to the smallest number of adjacent points exceeding that threshold to qualify as a cluster node; and (3) epsilon (ε), denoting the maximum distance within which points are considered neighbors. For K-means, we tested classification thresholds. These hyperparameters were selected because they govern the core mechanisms by which each algorithm defines spatial structures from the supervised model’s voxel-level probabilty output. DBSCAN relies on density and distance to detect contiguous high-risk regions of arbitrary shape, whereas K-means relies on a probability threshold to ensure only voxels with high predicted likelihood of progression influence centroid formation.

Table 1:

Results from hyperparameter tuning of DBSCAN and K-Means Clustering for the 14 patients included in the training and testing set of the underlying supervised Gradient Boosting classification model. Rows are sorted by AU-ROC; the peak model metrics (AUC: 0.942) were shown by DBSCAN at a 0.9 threshold, 2 minimum samples, and 1 EPS.

Model	Threshold	Minimum Samples	EPS	ROC-AUC	Accuracy	Sensitivity	Specificity	F1
DBSCAN	0.90	2	1	0.942	99.6%	88.5%	99.9%	91.5%
DBSCAN	0.99	3	2	0.940	99.5%	88.2%	99.8%	90.2%
DBSCAN	0.50	5	0.8	0.915	99.4%	83.2%	99.8%	88.4%
K-Means	0.50	-	-	0.874	98.9%	75.4%	99.4%	77.0%
K-Means	0.75	-	-	0.850	98.8%	70.5%	99.6%	74.8%
K-means	0.95	-	-	0.835	98.9%	67.3%	99.6%	75.3%

DBSCAN - Density-Based Spatial Clustering of Applications with Noise; AUC - area under the curve; AU-ROC - area under the receiver operating characteristic; EPS - epsilon parameter

The final clustering zone was used as a reference point for reclassifying our predicted AFR. All voxels predicted by the supervised Gradient Boosting model as AFR that were outside the unsupervised clustering zone were relabeled as “benign tissue”. False-negative voxels embedded within a core lesion cluster were relabeled as “high-risk” (Figure 2). An example case is shown in Figure 3

Figure 2:

Diagram of voxel re-classification after unsupervised clustering techniques. a) A mock axial MRI slice is visualized with supervised ML classifier predictions in each voxel, with values ranging from 0 (low likelihood for recurrence) to 1 (high likelihood for recurrence). b) After unsupervised clustering is performed, all voxels outside the target cluster are reclassified as healthy, non-recurrent tissue, while all voxels within the cluster are re-classified as areas of future recurrence (AFR) in a binary fashion. Created in https://BioRender.com.

Figure 3:

Example of WB-MRS unsupervised clustering predictions. A patient with stable disease on T1+CE (A), FLAIR (B), and fat-stranded T2 (C) post-treatment underwent WB-MRS imaging. Cho/NAA maps are overlaid (D) with the clustering zone retroactively superimposed for visualization. Using a supervised Gradient Boosting classifier, the likelihood of recurrence or progression in each voxel is plotted (E) along with unsupervised clustering (yellow). Surveillance contrast-enhanced imaging at Day 131 is shown (F) with medial and lateral expansion of the lesion.

1.2.4. Statistical Analysis

Based on reclassified predictions via the unsupervised cluster, we calculated sensitivity, specificity, accuracy, and F1-score to compare clustering-enhanced classifications to original, raw supervised model predictions. Receiver operating characteristic (ROC) curves were constructed, and AUC was computed. Metrics were calculated for the supervised model’s train-test cohort before being applied to the validation cohort.

1.3. Results

1.3.1. Unsupervised ML Hyperparameter Tuning

In total, 118,922 voxels were analyzed across all patients, with no missing metabolite data. After sucessfully reproducing the supervised Gradient Boosting tumor progression probabilities for each voxel, outputs were mapped in three-dimensional space for unsupervised clustering. The optimal model for DBSCAN was trained with a supervised Gradient Boosting progression threshold of 90%, minimum nearest neighbors of 2 and ε of 1. The optimal model for K-Means was trained on a threshold of 0.5, with a pre-initialized number of clusters at 2.

1.3.2. Unsupervised Clustering Performance

After hyperparameter tuning, DBSCAN achieved the highest overall performance among the tested methods. Compared to the baseline Gradient Boosting classifier (AUC: 0.86), DBSCAN had an AUC of 0.94 with improved accuracy (99.6%) and precision (94.6%). This indicates fewer false positives without compromising the model’s ability to capture tumor progression. Although DBSCAN’s sensitivity (88.5%) and F1-score (91.5%) were modestly lower than those of the supervised model alone (sensitivity: 98.5%; F1: 95.9%), DBSCAN showed increased specificity of 0.999. K-Means clustering offered a small improvement (AUC: 0.87) over baseline but did not surpass DBSCAN. A summary of hyperparameter tuning metrics and model outcomes is detailed in Table 1. Figure 3 provides axial slices demonstrating how clustering prunes tumor predictions that are distant from the main lesion core, and Figure 4 compares these overlays to clinically available raw WB-MRS overlays. Both Figures 3 & 4 show naked imaging from the day of WB-MRS as well as disease progression from surveillance imaging.

Figure 4:

Axial MRI and Whole Brain MR Spectroscopy (WB-MRS) with both supervised and unsupervised model overlays in the supervised training cohort as compared to clinical WB-MRS maps and tumor progression. This figure illustrates the application of supervised and unsupervised models in two training cohort patients. Contrast-enhanced T1-weighted MRI outlines disease at the time of WB-MRS (A) along with FLAIR (B) 57 days before progression on surveillance imaging (C). Comparatively, WB-MRS produced clinical Cho/NAA maps overlaid on fluid-attenuated inversion recovery (FLAIR) MRI, which are found in the medical record (D). The proposed visualization pipeline first reproduces these raw Cho/NAA maps (E) along with our overlaid unsupervised clustering target zone. Supervised gradient boosting classification is performed (F), with unsupervised clustering overlaid. The target cluster (yellow) is overlaid posteromedial to the lesion, where progressing contrast-enhancement and midline shift are present on surveillance imaging.

1.3.3. Validation Cohort Statistics and Visualization

Two patients were naïve to the underlying supervised Gradient Boosting model. When DBSCAN was applied to Gradient Boosting’s voxel-wise classifications, high accuracy was achieved at 97.4% and 95.8%, with 100% specificity in both models. Sensitivity was 27.4% and 11.3% respectively. Three sample slices from one validation patient are shown in Figure 5 as DICOM exports.

Figure 5:

Axial contrast-enhanced T1 MRI from a validation cohort patient overlaid with supervised model heatmaps and unsupervised clustering from WB-MRS data. Three unique slices (A-C, D-F, G-I) show supervised (red-blue) plus unsupervised (yellow) cluster predictions (A,D,G), clinical Cho/NAA maps overlaid on fluid attenuated inversion recovery (FLAIR) MRI (C,F,I). The re-registered Cho/NAA panels (B, E, H) served as a control for registration reproducibility of the clinical Cho/NAA maps. The third slice (G-I) is a control slice with no tumor.

1.4. Discussion

1.4.1. Overview

HGGs remain a persistent challenge in neuro-oncology due to their highly infiltrative nature and inevitable recurrence¹⁷. Conventional imaging modalities, particularly contrast-enhanced MRI, provide limited sensitivity in detecting early perilesional progression, necessitating the development of refined predictive tools¹⁸. In using WB-MRS to predict glioma progression, we demonstrate that an integrated pipeline with supervised voxel-wise ML⁵ and cluster-based unsupervised ML notably improves classification accuracy and spatial coherence. Our results indicate that DBSCAN clustering outperforms K-Means as the unsupervised ML model in this pipeline. These refinements yielded spatially consistent clusters that are combined with supervised ML heatmaps and overlaid onto T1-weighted contrast-enhanced MRI (Figure 4F). The results can be used right away as DICOM exports, highlighting its potential use in preoperative planning, intraoperative neuronavigation, or outpatient goals of care discussions. By incorporating spatial relationships between high-risk voxels, DBSCAN overcomes a key limitation of prior standalone supervised WB-MRS classification methods⁵, offering an organized visual representation of local tumor recurrence risk.

1.4.2. Unsupervised Clustering

Incorporating unsupervised clustering techniques to this pipeline has demonstrated potential in uncovering the true extent of tumor in radiographically occult regions. Our combined pipeline aims to detect compromised perilesional regions in HGGs via metabolites in the supervised ML phase and facilitate visualization of occult margins for tumor resection in the unsupervised clustering phase. This integration resulted in significant improvement in classification metrics from the supervised model alone. Key strengths include both a strong AUC and precision-recall balance.

The results underscore the advantage of density-based approaches in excluding voxel outliers and consolidating true tumor clusters, leading to spatially coherent margins for possible extent of diease delineation. In particular, DBSCAN demonstrated success in reclassifying false-positive voxel outliers that show inconsistent metabolic markers to its surroundings. The highly sensitive Gradient Boosting supervised model, while effective in identifying recurrence-prone areas, produced scattered false-positive predictions in healthy tissue (Figure 4F). Due to the density-based nature of DBSCAN, isolated voxels outside of high-risk tissue clusters were downgraded to progression-negative labels, converting high-resolution yet scattered voxel-wise predictions into a singular smoothened region of interest. This correction was particularly important in cases where isolated recurrence-prone voxels were initially binarized as positive despite model uncertainty. As a result, DBSCAN provided a more clinically actionable representation of tumor recurrence patterns.

A strong example is visible in Figure 4. The unsupervised clustering zone predicts posteromedial growth of the left frontal glioma. While the Cho/NAA maps alone do not fully capture the extent of metabolic changes in that region (Figure 4E), the surveillance imaging captures progressing contrast-enhancement along the falx cerebri with grossly visible midline shift (Figure 4C). By identifying clusters of high-risk voxels in direct proximity to resection margins, DBSCAN offers an additional piece of data to support clinical decision-making.

Within the validation cohort, DBSCAN shows highly conservative predictions with high specificity but low sensitivity. Given that supervised Gradient Boosting model is naive to the validation patients, its slightly weaker performance on new validation patients compared to the training cohort introduces confounding variation when assessing the additive value of the clustering models in a sequential pipeline. In the original validation of the supervised machining classifier, model metrics decreased compared to the training cohort; therefore, since these supervised predictions are the starting point for the unsupervised clustering algorithm, the lower starting accuracy may be contributing to the low sensitivity uncovered in this investigation.

Another possible explanation to the high specificity and lower sensitivity lies in the geometric constraints posed by DBSCAN. In one example, the predicted voxels of recurrence form a circumscribing polygonal space on the heatmap (Figure 5a) in three dimensions. A clustering algorithm like DBSCAN may struggle to track the contour of such a zone based on sparsely distributed voxels surrounding the high-risk zone, leading to under-segmentation as seen posteromedial to the cluster in Figure 5a. K-Means is at a greater risk to this vulnerability given that K-Means has a tendency to assign voxels to rigid centroids, instead of polygons, which may fail to capture the complex spatial heterogeneity of the tumor. Although K-Means consolidated high-risk voxels spatially, it was more prone to merging smaller distinct clusters or artificially splitting large clusters, potentially explaining its lower overall accuracy. Meanwhile, DBSCAN groups points based on density, automatically discovering clusters without pre-setting the cluster count.

Overall, both DBSCAN and K-Means provided valuable refinements of the initial supervised Gradient Boosting classification by incorporating spatial context, reducing false positive recurrence predictions, and improving the delineation of high-risk recurrence regions for utility in clinical segmentation. By refining spatial coherence, both models ensured that isolated “high-risk” voxels without proximity to a larger cluster were reclassified as healthy, reducing overdiagnosis and potential overtreatment. This increased specificity provides a more conservative calculation for the surgical resection ROI; joint use of the supervised and unsupervised predictions synergizes these specificity and sensitivity tradeoffs.

Our clustering-enhanced models exhibit higher specificity compared to the original supervised model, an important consideration for safety within surgical contexts where precision is critical. While the supervised model offers high sensitivity, its tendency to produce false positives could lead to unnecessary tissue resection. In contrast, the clustering models provide a conservative delineation of tumor margin, allowing for tailored resections and minimized risk of collateral damage to healthy parenchyma. Surgeons can utilize both models in tandem, leveraging the high sensitivity of the supervised model to identify potential areas of concern and the high specificity of the clustering models to refine high-risk tumor regions. With expanded datasets, these tools show promise in providing clinically actionable insights.

1.4.3. Neuronavigation Integration of the Resection Zone

The ability to integrate our clustering models into existing neuronavigation systems demonstrates translational potential by delineating tumor margins to guide a maximally safe resection. By exporting spatially coherent heatmaps as DICOM files, these models can be incorporated into existing neuronavigation platforms to provide surgeons with real-time, contoured visualization of high-risk recurrence zones. Of note, both the supervised scatter plot and unsupervised cluster zone are viewable (Figure 3E, 4F, 5A, 5D), thereby allowing surgeons to consider both the supervised heatmap as well as the unsupervised target cluster in context alongside functional tractography connectomics^19,20, perilesional vasculature²¹, and other more traditional modalities during preoperative planning.

1.4.4. Limitations

Our project is limited by multiple factors, mainly centered around sample size. The original cohort of patients during the previous supervised ML study was 16, with exposure to 14 during model creation and 2 saved for validation. This leaves the underlying supervised models vulnerable to overfitting, which occurs in smaller sample sizes whereby the model fits too closely to the underlying trained data and is less generalizable to a larger patient population. This may explain the differences between model performance metrics during the parameter optimization phase and performance on both validation patients. While the specificity of the cluster in the validation cohort was impressively high, the sensitivity remained low. This result is not unexpected, however, as the optimal hyperparameters used a threshold of 90% likelihood of recurrence to choose cluster nodes. As such, these hyperparameters serve as an advantage in forming a conservative surgical resection ROI but falls short in forming a ROI comprehensive to all relevant resection borders; future investigation of hyperparameters can investigate how hyperparameters may be tuned to accommodate clinical goals such as treatment aggression or quality of life optimization.

The supervised model was naïve to the validation patients and performed at a lower accuracy than it had with the training data, handicapping the results of the supervised to unsupervised model pipeline, speaking to the importance of larger datasets to optimize the training of both models. Both the supervised and unsupervised maps in tandem, as visualized in Figure 5, allow a surgeon to combine both approaches for a customized surgical approach, but this sensitivity tradeoff should be heavily considered in the further fine-tuning of this model with expanded datasets. The strength of this proof-of-concept study underscores the importance of expanding WB-MRS data collection to improve the generalizability of these recurrence prediction tools, in addition to further investigation of recurrence geometry.

Currently, the processing of raw spectrography data, application of ML progression prediction models, and exportation of clinical imaging maps is a stepwise process. Future developments should focus on automating the calculation and integration of WB-MRS progression prediction maps into neuronavigation software for a more seamless application of these tools. The creation of a pipeline automating this architecture from MRI data collection to prediction map availability into a Picture Archiving and Communication System (PACS) would greatly improve access to these models for clinical use, likened to the integration of automatic image fusion methods for co-registration of preoperative MRI and intraoperative ultrasound in the operative setting²².

Other limitations stem from the geometry specific to the clustering algorithms. Given that the training cohort’s highest AUC resulted from DBSCAN cluster parameters within 1 voxel, a model identifying one continuous cluster may fail to identify thin or hour-glass shaped contours. This possible constraint is visualized in Figure 5a, where the isthmus of recurrence voxels may prevent the cluster from continuing posteriorly.

1.5. Conclusion

The expansion of unsupervised clustering models coupled with supervised ML and WB-MRS data has the potential to both improve our understanding of tumor recurrence and change surgical resection strategies. This remains a proof-of-concept demonstration, and additional data collection—particularly with larger patient cohorts—will be necessary to validate and refine our approach. The translation of these clustering tools can create an ROI exportable into neuronavigation software for application in biopsy-guided research or expectant surgical management.

Highlights:

• High-grade glioma growth has been predicted by machine learning using metabolic data

• Unsupervised clustering of these predictions has been shown to further increase predictability

• Automated imaging pipelines can improve AI integration into neuro-oncologic workflow

Data Access:

The data presented in this study are available from the corresponding author, upon reasonable request and institutional approval. All proprietary python and R code is available upon reasonable request to the corresponding author.