From molecular signatures to radiomics: tailoring neurooncological strategies through forecasting of glioma growth

Philip Rauch Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Philip Rauch in
jns
Google Scholar
PubMed
Close
 MD
,
Martin Aichholzer Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Martin Aichholzer in
jns
Google Scholar
PubMed
Close
 MD
,
Carlo Serra Department of Neurosurgery, Clinical Neuroscience Center, University Hospital, University of Zurich, Switzerland;
Department of Neurosurgery, Machine Intelligence in Clinical Neuroscience (MICN) Lab, Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Switzerland;
Department of Neurology, Kepler University Hospital and Johannes Kepler University, Linz, Austria; and

Search for other papers by Carlo Serra in
jns
Google Scholar
PubMed
Close
 MD, PhD
,
Olivier Zanier Department of Neurosurgery, Machine Intelligence in Clinical Neuroscience (MICN) Lab, Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Switzerland;
Department of Neurology, Kepler University Hospital and Johannes Kepler University, Linz, Austria; and

Search for other papers by Olivier Zanier in
jns
Google Scholar
PubMed
Close
 BMed
,
Victor E. Staartjes Department of Neurosurgery, Clinical Neuroscience Center, University Hospital, University of Zurich, Switzerland;
Department of Neurosurgery, Machine Intelligence in Clinical Neuroscience (MICN) Lab, Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Switzerland;
Department of Neurology, Kepler University Hospital and Johannes Kepler University, Linz, Austria; and

Search for other papers by Victor E. Staartjes in
jns
Google Scholar
PubMed
Close
 MD, PhD
,
Petra Böhm Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Petra Böhm in
jns
Google Scholar
PubMed
Close
,
Gregor Seyer Institute of Statistics, Johannes Kepler University, Linz, Austria

Search for other papers by Gregor Seyer in
jns
Google Scholar
PubMed
Close
 BSc
,
Helga Wagner Institute of Statistics, Johannes Kepler University, Linz, Austria

Search for other papers by Helga Wagner in
jns
Google Scholar
PubMed
Close
 PhD
,
Ilja Manakov ImFusion GmbH, Munich, Germany;

Search for other papers by Ilja Manakov in
jns
Google Scholar
PubMed
Close
 PhD
,
Michael Sonnberger Institute of Neuroradiology, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Michael Sonnberger in
jns
Google Scholar
PubMed
Close
 MD
,
Nico Stroh Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Nico Stroh in
jns
Google Scholar
PubMed
Close
 MD
,
Stefan Aspalter Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Stefan Aspalter in
jns
Google Scholar
PubMed
Close
 MD
,
Kathrin Aufschnaiter-Hiessböck Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Kathrin Aufschnaiter-Hiessböck in
jns
Google Scholar
PubMed
Close
 MD
,
Tobias Rossmann Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Tobias Rossmann in
jns
Google Scholar
PubMed
Close
 MD
,
Annette Leibetseder Department of Neurology, Kepler University Hospital and Johannes Kepler University, Linz, Austria; and

Search for other papers by Annette Leibetseder in
jns
Google Scholar
PubMed
Close
 MD
,
Stefan Katletz Department of Neurology, Kepler University Hospital and Johannes Kepler University, Linz, Austria; and

Search for other papers by Stefan Katletz in
jns
Google Scholar
PubMed
Close
,
Andreas Gruber Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Andreas Gruber in
jns
Google Scholar
PubMed
Close
 MD, PhD
,
Matthias Gmeiner Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Matthias Gmeiner in
jns
Google Scholar
PubMed
Close
 MD, PhD
, and
Harald Stefanits Department of Neurosurgery, Kepler University Hospital and Johannes Kepler University, Linz, Austria;

Search for other papers by Harald Stefanits in
jns
Google Scholar
PubMed
Close
 MD, PhD
Free access

OBJECTIVE

Contemporary oncological paradigms for adjuvant treatment of low- and intermediate-grade gliomas are often guided by a limited array of parameters, overlooking the dynamic nature of the disease. The authors’ aim was to develop a comprehensive multivariate glioma growth model based on multicentric data, to facilitate more individualized therapeutic strategies.

METHODS

Random slope models with subject-specific random intercepts were fitted to a retrospective cohort of grade II and III gliomas from the database at Kepler University Hospital (n = 191) to predict future mean tumor diameters. Deep learning–based radiomics was used together with a comprehensive clinical dataset and evaluated on an external prospectively collected validation cohort from University Hospital Zurich (n = 9). Prediction quality was assessed via mean squared prediction error.

RESULTS

A mean squared prediction error of 0.58 cm for the external validation cohort was achieved, indicating very good prognostic value. The mean ± SD time to adjuvant therapy was 28.7 ± 43.3 months and 16.1 ± 14.6 months for the training and validation cohort, respectively, with a mean of 6.2 ± 5 and 3.6 ± 0.7, respectively, for number of observations. The observed mean tumor diameter per year was 0.38 cm (95% CI 0.25–0.51) for the training cohort, and 1.02 cm (95% CI 0.78–2.82) for the validation cohort. Glioma of the superior frontal gyrus showed a higher rate of tumor growth than insular glioma. Oligodendroglioma showed less pronounced growth, anaplastic astrocytoma—unlike anaplastic oligodendroglioma—was associated with faster tumor growth. Unlike the impact of extent of resection, isocitrate dehydrogenase (IDH) had negligible influence on tumor growth. Inclusion of radiomics variables significantly enhanced the prediction performance of the random slope model used.

CONCLUSIONS

The authors developed an advanced statistical model to predict tumor volumes both pre- and postoperatively, using comprehensive data prior to the initiation of adjuvant therapy. Using radiomics enhanced the precision of the prediction models. Whereas tumor extent of resection and topology emerged as influential factors in tumor growth, the IDH status did not. This study emphasizes the imperative of advanced computational methods in refining personalized low-grade glioma treatment, advocating a move beyond traditional paradigms.

ABBREVIATIONS

AIC = Akaike information criterion; EOR = extent of resection; F1 = superior frontal gyrus; IDH = isocitrate dehydrogenase; LGG = low-grade glioma; MSPE = mean squared prediction error; MTD = mean tumor diameter; OS = overall survival.

OBJECTIVE

Contemporary oncological paradigms for adjuvant treatment of low- and intermediate-grade gliomas are often guided by a limited array of parameters, overlooking the dynamic nature of the disease. The authors’ aim was to develop a comprehensive multivariate glioma growth model based on multicentric data, to facilitate more individualized therapeutic strategies.

METHODS

Random slope models with subject-specific random intercepts were fitted to a retrospective cohort of grade II and III gliomas from the database at Kepler University Hospital (n = 191) to predict future mean tumor diameters. Deep learning–based radiomics was used together with a comprehensive clinical dataset and evaluated on an external prospectively collected validation cohort from University Hospital Zurich (n = 9). Prediction quality was assessed via mean squared prediction error.

RESULTS

A mean squared prediction error of 0.58 cm for the external validation cohort was achieved, indicating very good prognostic value. The mean ± SD time to adjuvant therapy was 28.7 ± 43.3 months and 16.1 ± 14.6 months for the training and validation cohort, respectively, with a mean of 6.2 ± 5 and 3.6 ± 0.7, respectively, for number of observations. The observed mean tumor diameter per year was 0.38 cm (95% CI 0.25–0.51) for the training cohort, and 1.02 cm (95% CI 0.78–2.82) for the validation cohort. Glioma of the superior frontal gyrus showed a higher rate of tumor growth than insular glioma. Oligodendroglioma showed less pronounced growth, anaplastic astrocytoma—unlike anaplastic oligodendroglioma—was associated with faster tumor growth. Unlike the impact of extent of resection, isocitrate dehydrogenase (IDH) had negligible influence on tumor growth. Inclusion of radiomics variables significantly enhanced the prediction performance of the random slope model used.

CONCLUSIONS

The authors developed an advanced statistical model to predict tumor volumes both pre- and postoperatively, using comprehensive data prior to the initiation of adjuvant therapy. Using radiomics enhanced the precision of the prediction models. Whereas tumor extent of resection and topology emerged as influential factors in tumor growth, the IDH status did not. This study emphasizes the imperative of advanced computational methods in refining personalized low-grade glioma treatment, advocating a move beyond traditional paradigms.

Low-grade gliomas (LGGs) and intermediate-grade gliomas (herein subsumed as LGGs) are a distinct subset of neuro-oncological tumors that pose multifaceted challenges to both patients and clinicians. The inherent heterogeneity of LGGs, only superficially addressed by current pathological and molecular analyses, underscores the imperative need for more refined and patient-centric management strategies.1,2 Given the extended survival typically associated with LGGs, the primary objective has transcended mere life extension to encompass the assurance of optimal quality of life. This necessitates a consideration wherein repeated surgeries are more advantageous than early radiotherapy, which is known to potentially impair cognitive functions over time.35 By adopting a tailored approach specific to each patient’s glioma type, the possibility of neuro- or metaplasticity between surgeries emerges, which could be compromised if radiotherapy is pursued too early.6

In their quest for standardization and objectivity, neuro-oncologists and neurosurgeons typically strive to streamline their treatment approaches, aiming for comparability in adjuvant treatments and surgical timelines.7,8 This endeavor is not just for the benefit of future research but also to bolster argumentative rationales.

However, in contemporary neuro-oncological literature, the prevailing paradigms addressing the optimal initiation time frame for adjuvant therapy or the determination of the ideal surgical intervention window largely confine themselves to a limited set of parameters. These typically encompass genetic mutation signatures, patient age, and the extent of resection (EOR)—disregarding the potential significance of tumor topology, neuropsychological status, and tumor growth velocity.7,8 Such a standardized approach, which neglects the dynamic characteristics of the disease, might unintentionally lead to premature adjuvant interventions, thus jeopardizing the delicate balance between oncological outcomes and functional integrity on a patient-specific basis.9

Recent research underscores the importance of tumor growth velocity as a significant predictor of overall survival (OS) in LGGs.2,10,11 This dynamic parameter, determined through serial volumetric measurements, facilitates the early categorization of LGGs based on their proliferative kinetics. Such stratification holds potential for augmenting the precision in determining optimal intervention windows that are specifically tailored to address the multifaceted biological heterogeneity of these tumors. Basing predictions on molecular markers alone might oversimplify the intricate reality of LGG management. For instance, although isocitrate dehydrogenase (IDH) mutation status was shown to be a relevant predictor for OS, due to great heterogeneity across patients, its importance for predictions at an individual level remains questionable.12,13 This variability is further accentuated by factors like pregnancy or malignant transformation, which can significantly alter the natural course of LGGs.2,14 Given the dynamic nature of individual predictive factors, the challenge lies in effectively predicting tumor growth by using the full spectrum of available data.

Radiomics, a method characterized by its capability of objectively extracting features from radiographic images, has recently emerged as a very promising parameter subset with predictive value in this context.1518 By offering a detailed perspective of tumors, radiomics holds the potential to unveil the often overlooked characteristic of intratumoral heterogeneity,13,19,20 which is integral for determining tumor behavior, guiding therapeutic decisions, and forecasting overall prognosis. The emphasis on objective feature extraction becomes paramount for facilitating comparative analyses in future research endeavors.

In response to these challenges and opportunities, this study introduces a state-of-the-art prediction model for tumor growth velocity. By leveraging the potential of deep learning–based radiomics feature extraction in tandem with clinical patient data, this model aims to enhance the comprehension of LGG growth dynamics. This approach attempts to refine predictions of tumor behavior, presenting a more holistic methodology that merges sophisticated MRI analyses with molecular and comprehensive clinical data. Building on multicenter patient cohorts, our model offers a precise tool for predicting tumor volume throughout the disease’s progression, highlighting a significant advancement in truly personalized therapeutic approaches for LGGs.

Methods

Study Design and Study Cohort

The study was structured into 5 methodological stages, as illustrated in Fig. 1. The data were obtained from 2 independent cohorts with a total of 287 grade II and grade III gliomas.

FIG. 1.
FIG. 1.

Study workflow: a schematic representation of the research process, delineated into 5 distinct steps.

The pre- and postoperative MRI data, as well as clinical data for the training set were retrospectively collected between 2000 and 2023 from the LGG database of the Department of Neurosurgery, Kepler University Hospital (n = 278, 1760 observations). In the prospectively collected validation set, 9 patients were enrolled at the Department of Neurosurgery at the University Hospital Zurich. The clinicopathological information of both cohorts is summarized in Table 1. The surgeries were all performed as part of the clinical routine, with preoperative neuroradiological evaluation and serial postoperative follow-up until adjuvant therapy was initiated.

TABLE 1.

Comparative overview of relevant parameters between the training cohort from Kepler University Hospital and University Hospital Zurich

Clinical DataTraining Cohort—Kepler University Hospital Linz, n = 278Validation Cohort—University Hospital Zurich, n = 9
Age in yrs (range)36 (16–70)42 (25–58)
Sex
 Female1164
 Male1625
Mean no. of MRI observations (range)6.2 ± 5.0 (2–28)3.6 ± 0.7 (3–5)
≥3 MRI scans per patient (%)1116 (63.4%)9 (100%)
Time btwn 1st & last MRI sequences in days (range)1123 ± 1458 (27–6958)249 ± 290 (31–812)
No. of surgeries
 Biopsy210
 12036
 2453
 390
Initial tumor vol in cm³ (range)62.8 ± 59.7 (0–294.5)68.4 ± 53.3 (4.8–184.6)
MTD/yr in cm (95% CI)0.38 (0.25–0.51)1.02 (0.78–2.82)
Time to adjuvant therapy in mos (range)28.7 ± 43.3 (0–231)16.1 ± 14.6 (1–35)
Radiological malignancy at 1st MRI scan89 (32%)2 (22.2%)
EOR
 Gross-total resection133 (47.8%)2 (22.2%)
 Subtotal resection85 (30.6%)6 (66.7%)
 Biopsy60 (21.6%)1 (11.1%)
Mean resection rate, % (range)72.0 ± 27.0 (0–100)NA
WHO histological grade II1695
WHO histological grade III1094
Histological diagnosis
 Diffuse astrocytoma1082
 Anaplastic astrocytoma814
 Oligodendroglioma613
 Anaplastic oligodendroglioma280
IDH1 R132H/other IDH mutations188 (67.6%)9 (100%)
1p19 codeletion89 (32%)3 (33%)

NA = not available.

Unless otherwise indicated, values are expressed as the number of patients (%) or the mean ± SD.

The eligibility criteria comprised the following (for both cohorts): 1) histologically verified WHO grade II or III supratentorial glioma based on the WHO 2016 classification; 2) availability of preoperative MRI scans including T1-weighted with and without contrast enhancement, T2-weighted, and FLAIR; 3) at least 3 sequential follow-up MRI scans with minimum interval > 1 month before or after surgery without initiation of adjuvant chemotherapy or radiotherapy; and 4) no neoadjuvant treatment or previous cranial surgery. Patients were excluded for the following reasons: 1) if the automated deep learning segmentation showed erroneous results; 2) if artifacts or low image quality of the transversal slices of the aforementioned MRI sequence were likely to interfere with an acceptable radiomics analysis (i.e., susceptibility artifacts, restless patient); or 3) if there were intervals with volume decrease greater than 10% or mean tumor diameter (MTD) decrease greater than 2 mm due to resolving postoperative edema. Patients with missing data in relevant clinical variables (i.e., mutation status, histology) were excluded.

To fit the final prognostic models, 191 patients from Kepler University Hospital Linz were included. Nine patients from the Department of Neurosurgery, University Hospital Zurich, served as a validation cohort (Table 1).

Preoperative MRI scans underwent a qualitative assessment based on a modified topological and phylogenetic tumor extension protocol as proposed by Akeret et al.21,22 Anatomical structures considered merely displaced or edematous, without invasion, were categorized as unaffected. Occasionally, postresection MRI evaluations aided this determination.

Tumor morphology, based on its boundaries and the observable displacement of adjacent structures as well as their MRI signature, was classified as either expansive or diffuse. Differentiation was based on the tumors’ T1 signal characteristics. Compact tumors typically presented as hypointense on T1-weighted images, contrasting with their hyperintense appearance on T2-weighted images. In contrast, diffuse tumors were characterized by isointense signals on T1-weighted imaging. Contrast enhancement received a binary grade (yes/no), and tumors exhibiting contrast enhancement concurrent with an elevated regional cerebral blood volume were designated as malignant. Any subsequent significant increase of contrast enhancement, irrespective of regional cerebral blood volume alterations, was interpreted as malignant transformation in subsequent scans. Image evaluations were collaboratively executed by two experienced investigators (P.R. and M.A.), with a third expert (M.S.) mediating to achieve consensus if there were discrepancies.

Deep Learning Segmentation and Feature Extraction

We implemented a deep learning–based automatic segmentation process, complemented by automated feature extraction. An objective approach is essential for the radiomic evaluation of tumors, because accurate tumor segmentation masks are crucial and directly impact outcome parameters. To counteract potential variations arising from inter- and intrarater differences, we used a previously established U2–neural network, ensuring an unbiased collection of tumor segmentations.15 For the feature extraction we used the widely implemented PyRadiomics package.23 From an initial pool of 428 radiomics variables, we undertook a preselection based on the Pearson correlation as previously published.15

Volumetric Segmentation

We used deep learning–based segmentation for the initial MRI scans to ensure unbiased radiomics extraction. However, due to limited performance in certain postoperative MRI studies, follow-up analyses were conducted manually as described below. All eligible MRI studies underwent volumetric segmentation of T2/FLAIR anomalies at Kepler University Hospital. An expert neuroradiologist (M.S.) and neurosurgeon (P.R.) sequentially assessed the MRI datasets for each patient, spanning from the earliest to the most recent time points. All segmentations were executed using the semiautomated tool ImFusion Labels (ImFusion, 2022). An interactive level tracing tool facilitated the initial contouring of T2/FLAIR abnormalities. Subsequent manual edits to the segmented volume contour were made to incorporate tumor regions omitted in the initial contour or to exclude nontumor regions, such as postsurgical cavities, that had been inadvertently included in the preliminary contour.

Model Fitting

Random slope models with subject-specific random intercepts were fitted to the available values of the MTD of the included patients in the LGG cohort from Kepler University Hospital in two different scenarios: 1) based on preoperative information, and 2) using also available postoperative information.

For both models radiomics features were used. When radiomics variables showed a correlation exceeding 0.8, one from each pair was omitted, leaving a total of 99 distinct radiomics variables.

Our objective was to model MTD both pre- and postsurgery, prior to the initiation of adjuvant therapy, using all accessible data. By accounting for patient-specific heterogeneity in both initial size and growth, the model facilitates individualized predictions of MTD. The preoperative model, using only information available prior to the first surgery, aimed to gauge preoperative tumor growth rates. This insight could potentially inform the optimal timing for surgical intervention, with factors like tumor topology, size, and radiomics considered. The postoperative model, on the other hand, integrated a broader range of variables, including the EOR and molecular biology.

The response variable in the analyses was tumor volume, quantified using the formula 3√2 × Volume, where volume is expressed in cm3.10 In both models the available measurements of tumor volume were modeled depending on clinical and radiomic variables, while allowing for subject-specific heterogeneity with respect to tumor volume at the first measurement and its increase over time via random effects. The reference was a patient with a tumor situated in the superior frontal gyrus (F1) with oligodendroglial histology and baseline values in all other covariates (Table 2). The growthYear variable quantifies the average tumor growth in MTD per year for a patient with baseline values in all covariates.

TABLE 2.

Overview of postoperative baseline categories

VariableBaseline Category
GyrusF1
HistologyOligodendroglioma
IDHYes
1p19qNo
Surgery typeResection

Table details categories used as reference for the postoperative predictive model.

For the random slope models, an initial selection of diverse clinical variables served as dependent parameters. Following this, backward selection based on the Akaike information criterion (AIC) was used to streamline and identify the optimal model.

Finally, an external cohort of the Department of Neurosurgery, University Hospital Zurich, was used for model validation (n = 9). To evaluate prediction accuracy, we assumed that all MTD values before surgery and the first value after the first surgery of these patients are known, and predicted the next MTD at the time of the next measurement from the model. Given that estimation of subject-specific random effects requires at least three observations per patient, which would have further reduced our validation set, we set random intercepts and slopes for these patients to zero; i.e., predicted the next value using only covariate information of the patients, but did not exploit subject-specific deviations of tumor size and tumor growth. To assess prediction performance, we computed the mean squared prediction error (MSPE) per year as the squared difference between the MTD predicted from the model at the time of the next measurement, minus the observed value, divided by the length of the time interval from the first MTD after surgery (in years).

Ethics

Ethics board approval was obtained from the local ethics committee prior to data acquisition. All patients or their legal representatives gave their legal informed consent, and the study was conducted in accordance with the Declaration of Helsinki. The results are reported in accordance with the STROBE statement.

Code Availability

All statistical calculations were performed with R statistical software (version 4.3.1, 2022). The R code used for crafting the radiomic and prognostic models in this study can be accessed at https://github.com/PRauch1/LGG–velocity.git. However, the code associated with the framework, preprocessing, and segmentation is proprietary to ImFusion and is not publicly accessible. For further inquiries, the lead author is available to address questions upon reasonable request.

Results

Preoperative Model

This model enables prediction of future presurgical tumor growth using a single MRI scan. Key parameters included the specific gyrus affected, the degree of radiological malignancy, and whether the tumor exhibits a diffuse or expansive growth pattern, in conjunction with a detailed analysis of radiomics features.

It is important to emphasize that the covariate set of this model is inherently restricted, given that certain parameters, such as histology, can only be discerned postoperatively. As such, the presurgical model primarily offers a foundational framework for initial tumor growth projections. Only the gyrus and malignancy variables had a significant effect. Using forward selection, 11 of the 99 radiomics variables had significant effects and were integrated into our final model. For the purpose of this study, only in-sample validation could be performed, because too few measurements for reasonable out-of-sample validation were available (Fig. 2).

FIG. 2.
FIG. 2.

Observed MTD values and predictions from the preoperative glioma model with subject-specific random slopes and intercepts (in-sample, patients A–F). The blue line denotes observed volumetric measurements, whereas the red line depicts predicted tumor volumes (increments on the y-axis denote cm3).

Postoperative Model

Radiomics selection was performed as in the preoperative model and in-sample validation was also conducted (Fig. 3). Only significant variables from the preoperative model were considered. In Fig. 4, the variables labeled "postSurg" denote indicators of previous surgeries, whereas "resection 1–3" labels specify the proportion of tumor mass removed during the respective surgical procedures; different slopes were observed, depending on the EOR. Additional variables encompass attributes like tumor location (gyrus), histological classification, type of surgical intervention, and molecular markers associated with the tumor. Main effects are listed on the left, and interactions between tumor growth and specific tumor classifications—i.e., modification of the annual growth for these tumor types—are on the right. In our assessment of predictive models, we used the AIC as a measure of model performance. The model integrating radiomics data outperformed its clinical data-only counterpart, evidenced by a lower AIC value of 2991 compared to 3145. This superior model was then applied to an independent validation dataset to evaluate its out-of-sample predictive accuracy.

FIG. 3.
FIG. 3.

Observed MTD values and predictions from the postoperative model with subject-specific random slopes and intercepts (in-sample, patients A–F). The blue line represents actual volumetric measurements, whereas the red line shows tumor volume predictions. The red dotted lines mark surgeries 1–3. Increments on the y-axis denote cm3.

FIG. 4.
FIG. 4.

Confidence intervals for covariate effects with respect to baseline (tumor in F1, oligodendroglioma, IDH positive). The dotted line marks 0; confidence intervals at the left side indicate negative effects and at the right side indicate positive effects. Main effects, denoted by the name of the corresponding variable, indicate effects on the initial value of the glioma; "growthYears" denotes the yearly growth and, as a suffix, its modification due to a specific value of (for example) gyrus. The effect of "growthYears: postsurg: resection" variables shows the influence of EOR on tumor growth. wt = wild type.

Our findings indicate distinct variations based on tumor location and histology. Specifically, tumors in the precentral gyrus tended to be smaller than those in F1, and anaplastic oligodendrogliomas were generally larger than oligodendrogliomas at first visit.

Additionally, insular gliomas grew slower than those in the F1. With respect to histology, anaplastic astrocytomas, unlike anaplastic oligodendroglioma or WHO grade, showed a distinctively faster growth compared to oligodendrogliomas. Notably, there was a considerable reduction in tumor growth with increased EOR. These results emphasize the integral roles of tumor location, histological subtype, and resection in shaping tumor growth trajectories. Contrary to several radiomics variables depicted in Fig. 4, the effect of IDH on tumor growth was negligible. For illustrative purposes, we generated 2 in silico patients with high- and low-risk profiles and visualized their respective simulated MTDs per year (Fig. 5).

FIG. 5.
FIG. 5.

Comparative growth curve projections for distinct clinical scenarios. This figure contrasts the predicted postsurgical growth curves for 2 in silico patients: patient 1 (blue line) with oligodendroglioma (median radiomic features), and patient 2 (red line) with IDH-mutant anaplastic astrocytoma (75th percentile radiomic features), with all other variables at baseline. Increments on the y-axis denote cm3.

External Validation

In the out-of-sample predictions, 9 patients, who were not part of the primary dataset, were evaluated (Fig. 6). For these patients, data up to the first postsurgical observation were used to predict the subsequent tumor volume. As mentioned above, no random effects were estimated for these subjects, and both random intercepts and slopes were set to zero. Even without accounting for subject-specific random effects, predicted and actual values agree well—with the exception of 1 patient, whose tumor was an outlier (volume exceeding 170 cm3) and exhibited a highly diffuse character, which complicated accurate volumetric assessments. To quantify the prediction accuracy, the MSPE per year was computed across the remaining 8 patients, yielding an MSPE of 0.58 cm.

FIG. 6.
FIG. 6.

Predicted versus actual tumor volume for the external validation cohort (patients A–H). The blue curve traces the observed tumor volume over time, whereas the red curve corresponds to the predictions from the model. The point of surgical intervention is indicated by the dotted red line. Increments on the y-axis denote cm3.

Discussion

Forecasting the natural progression of brain tumors, especially within the diverse spectrum of LGGs, stands as a significant challenge in modern neuro-oncology.1,24 The inherent complexity and pronounced heterogeneity of potential outcomes can inadvertently lead to the adoption of oversimplified treatment paradigms.1,8 Such methodologies might not only dilute personalized patient care in the pursuit of methodological homogeneity but also precipitate untimely radiotherapy decisions anchored on a narrow spectrum of molecular markers that may provide limited value on an individual level.7 In this investigation, we aimed to overcome these limitations by harnessing a comprehensive dataset, integrating detailed anatomical information and precise and quantifiable resection metrics. An important aspect of our approach was the potential exploration of the often undervalued biological heterogeneity within LGGs,25,26 by using an objective deep learning–based radiomics methodology, together with an advanced multivariate prognostic model. This integrative approach emphasizes the critical role of sophisticated computational techniques in enhancing and personalizing therapeutic strategies for LGG management.

Our findings are consistent with existing literature, emphasizing the negligible impact of IDH mutation status on glioma growth rates.12 We further substantiate that surgical interventions do not inherently accelerate tumor growth;27 however, the EOR plays a pivotal role in modulating tumor progression. Distinctively, our study sheds light on the influence of tumor topology on growth dynamics, revealing a more indolent growth trajectory for tumors located in the insular region compared to those in the F1. Corroborating prior observations, oligodendrogliomas demonstrated a slower growth pattern,11 whereas anaplastic astrocytomas, but not anaplastic oligodendrogliomas or WHO grade, presented a more accelerated growth trajectory. A notable contribution of our investigation is the enhanced predictive accuracy achieved by incorporating radiomics variables, underscoring their potential utility in future oncological prognostic models.

In a multicenter validation setting, our model achieved an MSPE of 0.58 cm for anticipating future glioma volumes by using individualized patient data. It is imperative to acknowledge that predictive models, by their very nature, offer approximations. However, their clinical merit is amplified when they serve as robust empirical compasses guiding therapeutic decisions. For patients undergoing surgical intervention soon after diagnosis, the limited preoperative data can pose challenges in making informed clinical decisions. This often results in a reliance on conventional protocols that may not consider the tumor’s evolving nature. The complexity is accentuated when there is an immediate call for adjunctive therapies based on prevailing guidelines, especially in the absence of comprehensive postoperative imaging that remains unaffected by radiotherapy or chemotherapy interventions. Thus, by harnessing the granularity of advanced prognostic modeling, our findings point toward a nuanced shift in neuro-oncological practices. This refined approach promotes clearer, patient-focused decision-making, potentially facilitating discussions around strategies such as deferring radiotherapy in light of processes like metaplasticity,28 anchored in the tumor’s projected pathophysiological trajectory. Our study, while novel in several aspects, acknowledges certain inherent limitations. The retrospective training cohort, coupled with the limited cohort size in which more than 3 serial pre-/postoperative MRI scans were obtained prior to the initiation of adjuvant therapy, poses challenges. Notably, a significant subset of patients was excluded due to either suboptimal MRI quality for radiomics analysis or the absence of requisite clinical variables. This constraint led to broader confidence intervals, potentially affecting the robustness of our assessments.

A patient was excluded from the external validation cohort because of the model’s suboptimal performance in handling extreme outliers. This underscores the need for more comprehensive datasets to ensure precise modeling for exceptional cases. Additionally, the limited size of our validation group necessitates caution when generalizing our results to wider clinical contexts. Although we used the U2–neural network deep learning approach29 for tumor segmentation, the subsequent reliance on handcrafted radiomics features, derived from PyRadiomics, suggests potential room for enhancement. A transition to a deep learning–centric feature extraction might offer more nuanced insights.30

For the preoperative model, our dataset predominantly leaned on historical data from the wait-and-see era or the (rare) individual initial preferences of patients who opted for a wait-and-see approach. In the current context of prioritizing early and functionally maximal (or supramaximal) resection, the accuracy of preoperative volume estimations is constrained by the limited availability of sequential preoperative MRI sequences. Consequently, this study predominantly centers on postoperative predictions. Furthermore, the constraints of our dataset precluded a thorough exploration of supramaximal resection’s effects, despite its potential significance in glioma treatment dynamics.31

In future developments, our model is strategically designed to achieve greater robustness by integrating larger, multicenter datasets from high-volume centers for both training and validation. We also anticipate further advancements through the incorporation of a diverse range of diagnostic modalities, including PET scans, whole-genome sequencing, and advanced (functional) MRI techniques. The inclusion of more patients and advanced data sources is expected to further enhance the model’s predictive accuracy. Such a comprehensive integration is expected not only to bolster its clinical applicability but also to make a substantial contribution to the evolving realm of precision medicine.

Consequently, we advocate for the use of dynamic prediction models, because such models can support argumentative rationales,24 particularly in decisions surrounding the timing of radiotherapy initiation. By visualizing potential outcomes, patients are also better positioned to participate in their care decisions, optimizing a patient-centric onco-functional balance.

Our model offers a novel approach to glioma growth prediction by using radiomics analysis, supported by multicenter validation. Moving forward, discerning the diverse responses of gliomas to adjuvant treatments is also of high interest. Such insights will be important in enhancing and tailoring glioma simulations in future studies.

Conclusions

In this study we developed a patient-specific model to predict LGG volumes before chemo- and/or radiotherapy, underscoring the model’s potential for personalized treatment protocols. We assessed the intrinsic growth dynamics of LGG prior to any adjuvant intervention; multivariate analysis was performed using a wide spectrum of clinical parameters and deep learning–based radiomics. Our findings, corroborated by multicenter cohorts, underscore the enhanced predictive accuracy achieved by incorporating radiomics. Notably, whereas tumor EOR and topology emerged as influential factors on tumor growth, the IDH status did not. Our results underscore the pivotal role of sophisticated computational methodologies in shaping individualized LGG treatment strategies, advocating a move beyond conventional treatment paradigms.

Disclosures

The authors report no conflict of interest concerning the materials or methods used in this study or the findings specified in this paper.

Author Contributions

Conception and design: Aichholzer, Rauch, Wagner, Manakov, Leibetseder, Stefanits. Acquisition of data: Aichholzer, Rauch, Serra, Zanier, Staartjes, Böhm, Sonnberger, Stroh, Aspalter, Aufschnaiter-Hiessböck, Rossmann, Leibetseder, Gmeiner. Analysis and interpretation of data: Rauch, Seyer, Wagner, Stroh, Aspalter, Leibetseder, Katletz, Gmeiner. Drafting the article: Aichholzer, Rauch, Wagner, Manakov, Sonnberger, Leibetseder. Critically revising the article: Serra, Staartjes, Wagner, Sonnberger, Stroh, Aufschnaiter-Hiessböck, Rossmann, Leibetseder, Gruber, Gmeiner, Stefanits. Reviewed submitted version of manuscript: Rauch, Staartjes, Wagner, Stroh, Leibetseder, Gruber, Gmeiner, Stefanits. Approved the final version of the manuscript on behalf of all authors: Aichholzer. Statistical analysis: Seyer, Wagner, Katletz. Administrative/technical/material support: Staartjes, Böhm, Manakov, Sonnberger, Stroh. Study supervision: Gruber.

References

  • 1

    Duffau H. A Personalized longitudinal strategy in low-grade glioma patients: predicting oncological and neural interindividual variability and its changes over years to think one step ahead. J Pers Med. 2022;12(10):1621.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 2

    Pallud J, Blonski M, Mandonnet E, et al. Velocity of tumor spontaneous expansion predicts long-term outcomes for diffuse low-grade gliomas. Neuro Oncol. 2013;15(5):595606.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 3

    Duffau H. Why brain radiation therapy should take account of the individual structural and functional connectivity: toward an irradiation "à la carte". Crit Rev Oncol Hematol. 2020;154:103073.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 4

    Connor M, Karunamuni R, McDonald C, et al. Regional susceptibility to dose-dependent white matter damage after brain radiotherapy. Radiother Oncol. 2017;123(2):209217.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 5

    Zhu T, Chapman CH, Tsien C, et al. Effect of the maximum dose on white matter fiber bundles using longitudinal diffusion tensor imaging. Int J Radiat Oncol Biol Phys. 2016;96(3):696705.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 6

    Taillandier L, Obara T, Duffau H. What does quality of care mean in lower-grade glioma patients: a precision molecular-based management of the tumor or an individualized medicine centered on patient’s choices? Front Oncol. 2021;11:719014.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 7

    Weller M, van den Bent M, Preusser M, et al. EANO guidelines on the diagnosis and treatment of diffuse gliomas of adulthood. Nat Rev Clin Oncol. 2021;18(3):170186.

  • 8

    Buckner JC, Shaw EG, Pugh SL, et al. Radiation plus procarbazine, CCNU, and vincristine in low-grade glioma. N Engl J Med. 2016;374(14):13441355.

  • 9

    Duffau H. Paradoxes of evidence-based medicine in lower-grade glioma: to treat the tumor or the patient? Neurology. 2018;91(14):657662.

  • 10

    Pallud J, Taillandier L, Capelle L, et al. Quantitative morphological magnetic resonance imaging follow-up of low-grade glioma: a plea for systematic measurement of growth rates. Neurosurgery. 2012;71(3):729740.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 11

    Huang RY, Young RJ, Ellingson BM, et al. Volumetric analysis of IDH-mutant lower-grade glioma: a natural history study of tumor growth rates before and after treatment. Neuro Oncol. 2020;22(12):18221830.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 12

    Gozé C, Blonski M, Le Maistre G, et al. Imaging growth and isocitrate dehydrogenase 1 mutation are independent predictors for diffuse low-grade gliomas. Neuro Oncol. 2014;16(8):11001109.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 13

    Poulen G, Gozé C, Rigau V, Duffau H. Huge heterogeneity in survival in a subset of adult patients with resected, wild-type isocitrate dehydrogenase status, WHO grade II astrocytomas. J Neurosurg. 2018;130(4):12891298.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 14

    Pallud J, Mandonnet E, Deroulers C, et al. Pregnancy increases the growth rates of World Health Organization grade II gliomas. Ann Neurol. 2010;67(3):398404.

  • 15

    Rauch P, Stefanits H, Aichholzer M, et al. Deep learning-assisted radiomics facilitates multimodal prognostication for personalized treatment strategies in low-grade glioma. Sci Rep. 2023;13(1):9494.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 16

    Aerts HJWL, Velazquez ER, Leijenaar RTH, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun. 2014;5(1):4006.

  • 17

    Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749762.

  • 18

    Li G, Li L, Li Y, et al. An MRI radiomics approach to predict survival and tumour-infiltrating macrophages in gliomas. Brain. 2022;145(3):11511161.

  • 19

    Johnson B, Mazor T, Hong C, et al. GE-15 Clonal evolution and intratumoral heterogeneity of low-grade glioma genomes. Neuro Oncol. 2014;16(Suppl 5):v99.

  • 20

    Ryu YJ, Choi SH, Park SJ, Yun TJ, Kim JH, Sohn CH. Glioma: application of whole-tumor texture analysis of diffusion-weighted imaging for the evaluation of tumor heterogeneity. PLoS One. 2014;9(9):e108335.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 21

    Akeret K, Vasella F, Staartjes VE, et al. Anatomical phenotyping and staging of brain tumours. Brain. 2022;145(3):11621176.

  • 22

    Akeret K, Weller M, Krayenbühl N. The anatomy of neuroepithelial tumours. Brain. 2023;146(8):31333145.

  • 23

    Welcome to pyradiomics documentation! pyradiomics v3.1.0rc2.post5+g6a761c4 documentation. Pyradiomics. Accessed December 18, 2023. https://pyradiomics.readthedocs.io/en/latest/

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 24

    Zhao R, Krauze AV. Survival prediction in gliomas: current state and novel approaches. In: Debinski W, ed. Gliomas. Exon Publications;2021.Accessed December 18, 2023. http://www.ncbi.nlm.nih.gov/books/NBK570711/

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 25

    Darlix A, Rigau V, Fraisse J, Gozé C, Fabbro M, Duffau H. Postoperative follow-up for selected diffuse low-grade gliomas with WHO grade III/IV foci. Neurology. 2020;94(8):e830e841.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 26

    Roodakker KR, Alhuseinalkhudhur A, Al-Jaff M, et al. Region-by-region analysis of PET, MRI, and histology in en bloc-resected oligodendrogliomas reveals intra-tumoral heterogeneity. Eur J Nucl Med Mol Imaging. 2019;46(3):569579.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 27

    Mandonnet E, Pallud J, Fontaine D, et al. Inter- and intrapatients comparison of WHO grade II glioma kinetics before and after surgical resection. Neurosurg Rev. 2010;33(1):9196.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 28

    Duffau H. Introducing the concept of brain metaplasticity in glioma: how to reorient the pattern of neural reconfiguration to optimize the therapeutic strategy. J Neurosurg. 2021;136(2):613617.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 29

    Qin X, Zhang Z, Huang C, Dehghan M, Zaiane OR, Jagersand M. U2-Net: going deeper with nested U-structure for salient object detection. Pattern Recognit. 2020;106:107404.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 30

    Han W, Qin L, Bay C, et al. Deep transfer learning and radiomics feature prediction of survival of patients with high-grade gliomas. AJNR Am J Neuroradiol. 2020;41(1):4048.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 31

    de Leeuw CN, Vogelbaum MA. Supratotal resection in glioma: a systematic review. Neuro Oncol. 2019;21(2):179188.

  • Collapse
  • Expand
  • FIG. 1.

    Study workflow: a schematic representation of the research process, delineated into 5 distinct steps.

  • FIG. 2.

    Observed MTD values and predictions from the preoperative glioma model with subject-specific random slopes and intercepts (in-sample, patients A–F). The blue line denotes observed volumetric measurements, whereas the red line depicts predicted tumor volumes (increments on the y-axis denote cm3).

  • FIG. 3.

    Observed MTD values and predictions from the postoperative model with subject-specific random slopes and intercepts (in-sample, patients A–F). The blue line represents actual volumetric measurements, whereas the red line shows tumor volume predictions. The red dotted lines mark surgeries 1–3. Increments on the y-axis denote cm3.

  • FIG. 4.

    Confidence intervals for covariate effects with respect to baseline (tumor in F1, oligodendroglioma, IDH positive). The dotted line marks 0; confidence intervals at the left side indicate negative effects and at the right side indicate positive effects. Main effects, denoted by the name of the corresponding variable, indicate effects on the initial value of the glioma; "growthYears" denotes the yearly growth and, as a suffix, its modification due to a specific value of (for example) gyrus. The effect of "growthYears: postsurg: resection" variables shows the influence of EOR on tumor growth. wt = wild type.

  • FIG. 5.

    Comparative growth curve projections for distinct clinical scenarios. This figure contrasts the predicted postsurgical growth curves for 2 in silico patients: patient 1 (blue line) with oligodendroglioma (median radiomic features), and patient 2 (red line) with IDH-mutant anaplastic astrocytoma (75th percentile radiomic features), with all other variables at baseline. Increments on the y-axis denote cm3.

  • FIG. 6.

    Predicted versus actual tumor volume for the external validation cohort (patients A–H). The blue curve traces the observed tumor volume over time, whereas the red curve corresponds to the predictions from the model. The point of surgical intervention is indicated by the dotted red line. Increments on the y-axis denote cm3.

  • 1

    Duffau H. A Personalized longitudinal strategy in low-grade glioma patients: predicting oncological and neural interindividual variability and its changes over years to think one step ahead. J Pers Med. 2022;12(10):1621.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 2

    Pallud J, Blonski M, Mandonnet E, et al. Velocity of tumor spontaneous expansion predicts long-term outcomes for diffuse low-grade gliomas. Neuro Oncol. 2013;15(5):595606.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 3

    Duffau H. Why brain radiation therapy should take account of the individual structural and functional connectivity: toward an irradiation "à la carte". Crit Rev Oncol Hematol. 2020;154:103073.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 4

    Connor M, Karunamuni R, McDonald C, et al. Regional susceptibility to dose-dependent white matter damage after brain radiotherapy. Radiother Oncol. 2017;123(2):209217.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 5

    Zhu T, Chapman CH, Tsien C, et al. Effect of the maximum dose on white matter fiber bundles using longitudinal diffusion tensor imaging. Int J Radiat Oncol Biol Phys. 2016;96(3):696705.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 6

    Taillandier L, Obara T, Duffau H. What does quality of care mean in lower-grade glioma patients: a precision molecular-based management of the tumor or an individualized medicine centered on patient’s choices? Front Oncol. 2021;11:719014.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 7

    Weller M, van den Bent M, Preusser M, et al. EANO guidelines on the diagnosis and treatment of diffuse gliomas of adulthood. Nat Rev Clin Oncol. 2021;18(3):170186.

  • 8

    Buckner JC, Shaw EG, Pugh SL, et al. Radiation plus procarbazine, CCNU, and vincristine in low-grade glioma. N Engl J Med. 2016;374(14):13441355.

  • 9

    Duffau H. Paradoxes of evidence-based medicine in lower-grade glioma: to treat the tumor or the patient? Neurology. 2018;91(14):657662.

  • 10

    Pallud J, Taillandier L, Capelle L, et al. Quantitative morphological magnetic resonance imaging follow-up of low-grade glioma: a plea for systematic measurement of growth rates. Neurosurgery. 2012;71(3):729740.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 11

    Huang RY, Young RJ, Ellingson BM, et al. Volumetric analysis of IDH-mutant lower-grade glioma: a natural history study of tumor growth rates before and after treatment. Neuro Oncol. 2020;22(12):18221830.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 12

    Gozé C, Blonski M, Le Maistre G, et al. Imaging growth and isocitrate dehydrogenase 1 mutation are independent predictors for diffuse low-grade gliomas. Neuro Oncol. 2014;16(8):11001109.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 13

    Poulen G, Gozé C, Rigau V, Duffau H. Huge heterogeneity in survival in a subset of adult patients with resected, wild-type isocitrate dehydrogenase status, WHO grade II astrocytomas. J Neurosurg. 2018;130(4):12891298.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 14

    Pallud J, Mandonnet E, Deroulers C, et al. Pregnancy increases the growth rates of World Health Organization grade II gliomas. Ann Neurol. 2010;67(3):398404.

  • 15

    Rauch P, Stefanits H, Aichholzer M, et al. Deep learning-assisted radiomics facilitates multimodal prognostication for personalized treatment strategies in low-grade glioma. Sci Rep. 2023;13(1):9494.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 16

    Aerts HJWL, Velazquez ER, Leijenaar RTH, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun. 2014;5(1):4006.

  • 17

    Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749762.

  • 18

    Li G, Li L, Li Y, et al. An MRI radiomics approach to predict survival and tumour-infiltrating macrophages in gliomas. Brain. 2022;145(3):11511161.

  • 19

    Johnson B, Mazor T, Hong C, et al. GE-15 Clonal evolution and intratumoral heterogeneity of low-grade glioma genomes. Neuro Oncol. 2014;16(Suppl 5):v99.

  • 20

    Ryu YJ, Choi SH, Park SJ, Yun TJ, Kim JH, Sohn CH. Glioma: application of whole-tumor texture analysis of diffusion-weighted imaging for the evaluation of tumor heterogeneity. PLoS One. 2014;9(9):e108335.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 21

    Akeret K, Vasella F, Staartjes VE, et al. Anatomical phenotyping and staging of brain tumours. Brain. 2022;145(3):11621176.

  • 22

    Akeret K, Weller M, Krayenbühl N. The anatomy of neuroepithelial tumours. Brain. 2023;146(8):31333145.

  • 23

    Welcome to pyradiomics documentation! pyradiomics v3.1.0rc2.post5+g6a761c4 documentation. Pyradiomics. Accessed December 18, 2023. https://pyradiomics.readthedocs.io/en/latest/

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 24

    Zhao R, Krauze AV. Survival prediction in gliomas: current state and novel approaches. In: Debinski W, ed. Gliomas. Exon Publications;2021.Accessed December 18, 2023. http://www.ncbi.nlm.nih.gov/books/NBK570711/

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 25

    Darlix A, Rigau V, Fraisse J, Gozé C, Fabbro M, Duffau H. Postoperative follow-up for selected diffuse low-grade gliomas with WHO grade III/IV foci. Neurology. 2020;94(8):e830e841.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 26

    Roodakker KR, Alhuseinalkhudhur A, Al-Jaff M, et al. Region-by-region analysis of PET, MRI, and histology in en bloc-resected oligodendrogliomas reveals intra-tumoral heterogeneity. Eur J Nucl Med Mol Imaging. 2019;46(3):569579.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 27

    Mandonnet E, Pallud J, Fontaine D, et al. Inter- and intrapatients comparison of WHO grade II glioma kinetics before and after surgical resection. Neurosurg Rev. 2010;33(1):9196.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 28

    Duffau H. Introducing the concept of brain metaplasticity in glioma: how to reorient the pattern of neural reconfiguration to optimize the therapeutic strategy. J Neurosurg. 2021;136(2):613617.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 29

    Qin X, Zhang Z, Huang C, Dehghan M, Zaiane OR, Jagersand M. U2-Net: going deeper with nested U-structure for salient object detection. Pattern Recognit. 2020;106:107404.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 30

    Han W, Qin L, Bay C, et al. Deep transfer learning and radiomics feature prediction of survival of patients with high-grade gliomas. AJNR Am J Neuroradiol. 2020;41(1):4048.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 31

    de Leeuw CN, Vogelbaum MA. Supratotal resection in glioma: a systematic review. Neuro Oncol. 2019;21(2):179188.

Metrics

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 1349 1349 252
PDF Downloads 937 937 62
EPUB Downloads 0 0 0