This retrospective study included 288 patients with newly diagnosed diffuse gliomas who underwent surgical resection at Tokyo Women’s Medical University Hospital between 2016 and 2024. All tumors were classified according to the WHO 2021 CNS tumor classification system.

The IDH1 mutation status was initially evaluated using immunohistochemistry (IHC), and cases negative for IDH1-R132H were further analyzed for both IDH1 and IDH2 mutations using direct sequencing. 1p/19q co-deletion status was assessed using either fluorescence in situ hybridization or multiplex ligation-dependent probe amplification. Based on integrated histological and molecular diagnoses, the dataset included 141 IDH-wild-type astrocytomas (Astro-WD), 68 IDH-mutant astrocytomas (Astro-MT), and 79 oligodendrogliomas (Oligo). All patients underwent iFC analysis of tumor samples obtained during surgery.

Preoperative neuroimaging features were extracted from MRI, CT, and ¹¹C-Met-PET scans, including the presence or absence of a T2-FLAIR mismatch, gadolinium enhancement, calcification on either T2*-weighted MRI or CT, and tumor-to-normal uptake ratio (TNR) on MET PET. Patient age was also included as a clinical variable. Each imaging feature was independently evaluated by three board-certified neurosurgeons and a majority vote was used to finalize the label.

T2-FLAIR mismatch, enhancement, and calcification were assessed using a standardized three-point scale: “+” (clearly present), “+/−” (ambiguous or mild), and “−” (clearly absent). Intraoperative tumor samples were processed using iFC to generate nuclear DNA histograms (binned into 10 intervals from the original 512-bin signal), detect aneuploidy, and compute the MI. Continuous variables (age, TNR, and MI) were normalized using Z-score standardization, and categorical features (four imaging variables and aneuploidy) were one-hot encoded. All features were then concatenated into a single fused vector as input to the machine-learning models.

The classification task aimed to predict molecular subtypes across three classes: Astro-WD, Astro-MT, and Oligo. We employed a Random Forest as the benchmark model, with key hyperparameters such as the number of estimators and the splitting criterion optimized within each cross-validation fold. Model performance was evaluated using stratified five-fold cross-validation with a fixed random seed for reproducibility (20). As a deep learning comparator, we implemented TabNet, a state-of-the-art interpretable model optimized for structured and tabular data (21). TabNet leverages sparse attention mechanisms to dynamically select relevant features at each decision step.

Training was conducted using a batch size of 128 and 10 decision steps, and an AdamW optimizer with early stopping (patience = 10 epochs) to avoid overfitting. The attention-based feature selection process in TabNet also enables the visual interpretability of predictive signals (22). Classification performance was evaluated using accuracy, F₁-score, and receiver operating characteristic (ROC)-AUC. Statistical comparison between models was performed using the DeLong test. Feature importance was assessed using the permutation importance for Random Forest and attention weights for TabNet (23). An overview of the data processing pipeline, including imaging and iFC-derived features and model architectures, is shown in Figure 1.

This study was approved by the Institutional Review Board of Tokyo Women’s Medical University (Approval No. 3540-R6). Written informed consent was obtained from all participants in accordance with the institutional and national guidelines.

We developed classification models for three molecular subtypes—Astro-WD, Astro-MT, and Oligo—using integrated features derived from preoperative imaging (gadolinium enhancement, T2-FLAIR mismatch, calcification on T2*-weighted MRI or CT, and TNR on ¹¹C-MET PET), iFC (DNA histogram, aneuploidy status, MI), and patient age.

To ensure the reliability of radiological annotations, inter-rater agreement among the three neurosurgeons was assessed using weighted Cohen’s kappa statistics. Substantial agreement was observed for gadolinium enhancement (κ = 0.859) and calcification (κ = 0.717), while T2 hypointensity (κ = 0.485) and the T2-FLAIR mismatch sign (κ = 0.452) showed moderate agreement. These results confirm that the imaging features used for model input were reproducible and clinically reliable.

In the dataset of 288 cases, the Random Forest model achieved an overall classification accuracy of 76.0%, with a macro-average ROC-AUC of 0.88 and micro-average ROC-AUC of 0.89 (Figure 2). The F₁-scores were 0.82 for Astro-WD, 0.72 for Astro-MT, and 0.68 for Oligo, indicating favorable classification performance across all three subtypes. The sensitivity and specificity were also stable, as shown in Table 1 and the confusion matrix in Table 2.

The macro-average ROC-AUC represents the unweighted mean of the AUC values across all classes, offering a balanced measure of model performance. By contrast, the micro-average ROC-AUC was calculated from pooled predictions and reflected class imbalance, making it more representative of real-world performance. These values are summarized in Figure 2.

To evaluate the contribution of each modality, we trained models using limited feature sets. The imaging + age model achieved 66% accuracy, while the iFC + age model reached 72%. The full model combining imaging, iFC, and age achieved 76%. Although the imaging + age model (66%) underperformed compared to imaging alone (69%), this outcome should not be interpreted as evidence that adding age universally degrades imaging-based prediction performance. It likely reflects interactions specific to the current model or dataset configuration. Notably, the addition of iFC data improved the overall classification accuracy by approximately 7 percentage points compared to imaging features alone. These findings highlight the complementary roles of imaging and iFC features, as well as the additive value of age. Confusion matrices are provided in Supplementary Table 1.

The feature importance was evaluated using the Gini importance (mean decrease in impurities) calculated during Random Forest training (Figure 3). The most influential feature was the TNR on MET-PET, followed by the MI, age, DNA histogram bin 6, and gadolinium enhancement. These features are thought to reflect the key biological properties associated with tumor metabolism, cellular proliferation, patient characteristics, and vascular permeability.

In particular, bin 6 of the DNA histogram made a substantial contribution, highlighting the utility of histogram-derived features in predicting molecular subtypes. This suggests that the nuclear DNA distribution data obtained from iFC provide valuable information beyond that captured by imaging alone.

Beyond Random Forest, we also examined feature importance using TabNet (Supplementary Figure 1). Notably, the top-ranked features—histogram6, TNR, patient age, T2-FLAIR mismatch, and gadolinium enhancement—were largely consistent between both classifiers, indicating a stable feature selection process. This agreement reinforces the reliability of these features across different interpretable models.

Furthermore, a comparison between correctly classified and misclassified cases revealed significant differences in TNR, MI, and age for specific subtypes (Figures 4A–F and Table 3). For example, TNR and age differed significantly in Astro-WD, whereas MI was significantly different in both Astro-MT and Oligo. These results indicate that the model relies on clinically meaningful variables to guide classification.

Among the three molecular subtypes, oligodendroglioma demonstrated the lowest F₁-score (0.68), with 22 out of 77 cases misclassified—primarily as Astro-WD, and to a lesser extent as Astro-MT (Table 2). Misclassified Oligo cases exhibited significantly higher MET-PET uptake (median TNR = 3.07 vs. 2.03, p = 0.0026) and malignancy index (11.9 vs. 4.67, p = 0.018) compared to correctly classified cases, while patient age did not differ significantly (p = 0.396) (Table 3). These findings suggest that some Oligo tumors may share biological features with Astro-WD, such as increased metabolic activity or abnormal nuclear DNA content, making classification more challenging. Further refinement of the model or inclusion of additional molecular markers may be required to improve prediction accuracy in such cases.

We developed a prototype real-time prediction application based on the Random Forest model for intraoperative support. The application used preoperative imaging and intraoperative iFC data as inputs and outputs molecular subtype predictions in the form of probability distributions, which were visualized as pie charts for intuitive interpretation.

This tool enabled real-time prediction of the genetic subtype during surgery and supported intraoperative decision-making, such as determining the EOR. The user interface was designed to be intuitive and informative for both surgeons and support staff, allowing quick comprehension of the predictive results.

The model inference was completed in less than one second on a Microsoft Surface Go 2 (Intel Core m3-8100Y, 8GB RAM), resulting in negligible latency. In the current workflow, radiological features are assessed and entered preoperatively, and intraoperative flow cytometry results are available approximately 10 min after tissue sampling. Data transfer is performed via SD card, although direct network integration is under consideration. Taking into account data entry and transfer time, the entire process from sampling to prediction output typically takes about 15 min, which is well within the intraoperative decision-making window.

The novelty of our model lies in the integration of quantitative features derived from iFC, such as the MI, aneuploidy, and DNA histograms, with conventional imaging data for molecular classification. Although iFC has traditionally been used for rapid intraoperative assessments of malignancy and tumor typing, our approach demonstrates that these data can be quantitatively incorporated into machine learning algorithms to enhance diagnostic precision beyond that provided by imaging alone (15–17, 24, 25).

Our previous study proposed an intraoperative surgical strategy based on the molecular data obtained during surgery (18). This study expands on this concept by providing an AI-assisted framework for real-time predictions. Recent advances in intraoperative genetic testing, including real-time and digital droplet polymerase chain reaction, have highlighted the clinical utility of intraoperative molecular diagnostics (26, 27).

Feature importance analysis based on Gini impurities revealed that the TNR on ¹¹C-MET PET, MI, patient age, bin 6 from DNA histograms, and gadolinium enhancement were the most influential predictors (Figure 3). The TNR, a known indicator of tumor metabolic activity, is associated with tumor grade and malignancy (28). MI reflects abnormalities in DNA content and has been linked to tumor aggressiveness and prognosis (15–18). Age has also been reported as a prognostic factor, particularly in lower-grade astrocytomas (28), and was notably lower in misclassified Astro-WD cases in our analysis.

The substantial contribution of histogram-derived features—especially bin 6—suggests that nuclear DNA distribution profiles captured by iFC contain valuable information that is not readily apparent from imaging alone.

In addition, comparisons between correctly and incorrectly classified cases showed statistically significant differences in the T/N ratio, MI, and age for certain subtypes (Figures 4A–F and Table 3). For instance, Astro-WD misclassification was associated with lower TNR and younger age, whereas Astro-MT and Oligo misclassifications were related to MI values, indicating that these variables were central to the model’s decision-making process.

This agreement reinforces the reliability of these features across different interpretable models.

To contextualize our model choice, we additionally evaluated LightGBM and XGBoost, which achieved accuracies of 0.72 and 0.73, respectively. These values were comparable to our primary models, while a baseline SVM classifier showed lower accuracy (0.64). In summary, Random Forest demonstrated the highest accuracy in our study. This finding implies that Random Forest may offer enhanced robustness to overfitting relative to gradient boosting and deep learning approaches, as well as improved adaptability to multiclass and imbalanced classification tasks compared to SVM.

Among the three molecular subtypes, the model performed least accurately for Oligo, with an F₁-score of 0.68. This may be partly due to the overlap in imaging characteristics, such as elevated ¹¹C-MET PET uptake, with Astro-WD. The iFC-derived features also showed limited specificity in a subset of Oligo cases.

Nevertheless, in illustrative case 2, the model accurately predicted an oligodendroglioma despite the initial suspicion as IDH wild-type based on imaging and clinical impression. This highlights the utility of iFC data as a complementary input when imaging-based assessments are inconclusive.

This study extended our previous research on iFC for intraoperative grading and tumor classification by demonstrating its applicability to molecular subtyping. The development of a functional application based on our model represents a meaningful step toward real-time intraoperative implementation.

Future integration with deep learning models and radiomics features derived from preoperative imaging could further enhance the diagnostic performance (29, 30). Our model may serve as a benchmark for the next generation of intraoperative decision-support systems for glioma surgery.

This retrospective study was conducted at a single institution. Thus, external validation and further case accumulation are essential to ensure generalizability. Some misclassified cases show atypical imaging or iFC patterns, underscoring the potential need for new markers or additional molecular data in select cases.

In the future, the integration of real-time quantitative data, including iFC and intraoperative genetic diagnostics, may enable the prediction of not only molecular subtypes, but also tumor grading, further supporting precise surgical decision-making.

Finally, we note that this study focused solely on molecular subtype classification and did not evaluate glioma grading (i.e., low vs. high grade) using ROC analysis. Although the malignancy index has been associated with tumor grade in prior work (16), further studies will be needed to extend this model to real-time grading support.