Clinical outcome and deep learning imaging characteristics of patients treated by radio-chemotherapy for a “molecular” glioblastoma

Abstract

Background

Since 2021, glioblastomas have been classified into two subgroups: classic glioblastomas (histGB), defined as IDH wild-type grade 4 astrocytomas with necrosis and vascular proliferation, showing contrast enhancement (CE) on MRI; and molecular glioblastomas (molGB), characterized by specific alterations (7+/10−, EGFR amplification, TERT mutation). Although not always the case, molGB often lack CE and may mimic low-grade gliomas (LGG), hence complicating the diagnosis. Survival outcomes remain debated. This study aimed to evaluate the response of molGB to standard treatment and assess the ability of machine learning and deep learning to differentiate molGB without CE from LGG on MRI.

Methods

We retrospectively studied 132 glioblastoma patients treated with radiotherapy and temozolomide, comparing the survival outcomes of histGB and molGB. Artificial intelligence (AI) models were trained using features from MRI FLAIR hypersignal segmentation to distinguish molGB without CE from LGG.

Results

No significant difference in median overall survival (OS) (20.6 vs 18.4 months, P = .2) or progression-free survival (10.1 vs 9.3 months, P = .183) was observed between molGB and histGB. However, molGB without CE demonstrated improved median OS (31.2 vs 18 months, hazard ratios 0.45). Artificial intelligence models distinguished molGB without CE from LGG, achieving a best-performing ROC AUC of 0.85.

Conclusions

While patients with molGB and histGB have similar overall survival, patients with molGB without CE appear to have better outcomes. Artificial intelligence models effectively differentiate molGB from LGG, supporting their potential diagnostic utility.

Implications for Practice.

Molecular glioblastomas (molGB) are grade 2 or 3 astrocytomas with molecular characteristics which, despite lacking necrosis or microvascular proliferation, are now classified as glioblastoma and treated likewise. They often show distinct radiological features like T2-FLAIR hypersignal without contrast enhancement on MRI, thus mimicking low-grade lesions. We show that patients with noncontrast-enhancing molGB exhibit significantly increased survival compared to those with other glioblastomas, when treated by surgery and standard radio-chemotherapy treatment. Moreover, deep learning may aid diagnosis owing to the noncontrast characteristic of molGB. This study enhances understanding of this glioblastoma subgroup and their response to standard treatment.

Introduction

Glioblastoma is the most common and aggressive primary brain tumor in adults. The prognosis remains very poor, despite standard treatment consisting in resection as complete as possible, followed by concomitant radiotherapy with daily temozolomide, continued in maintenance, with the recent option of Tumor-Treating Fields. Median progression-free survival (PFS) and median overall survival (OS) are thus increased to 6.7 and 20.9 months, respectively.^1-3 The cIMPACT NOW publication in 2018 reviewing the heterogeneous group of IDH wild-type (wt) diffuse gliomas WHO grade 2 and 3 established molecular criteria for identifying glioblastoma-like tumors, temporarily named molecular glioblastomas (molGB). The presence of a TERT promoter (TERTp) mutation, EGFR amplification, or gain of chromosome 7 combined with loss of chromosome 10 (chr +7/−10) in IDH wt astrocytoma is currently sufficient to classify the tumor as a grade 4, subgrouped under the term “glioblastoma IDH wt,” even without the usual histological criteria such as tumor necrosis or microvascular proliferation (Figure 1). Thus, they benefit from the same diagnostic recommendations, management, and follow-up as other glioblastomas.^4-9 Very few studies have examined the survival of these patients; the data currently available are controversial, and it remains unclear whether their prognosis is indeed similar to that of other glioblastomas with WHO grade 4 histological criteria (histGB). To our knowledge, no study to date has compared survival data for molecular glioblastoma versus conventional glioblastoma since they receive the same standard treatment.

Figure 1.

“Molecular glioblastoma” in the WHO classification of gliomas. Histologic Glioblastoma (Hist) refers to classic glioblastomas defined as IDH wild-type grade 4 astrocytomas with necrosis and vascular proliferation. “Molecular Glioblastoma” refers to IDH wild-type diffuse gliomas WHO grade 2 and 3 that meet molecular criteria making them glioblastoma-like tumors, thus classifying them as grade 4, with or without gadolinium enhancement, resembling either low- or high-grade tumors on MRI.

Glioblastomas are typically identified on imaging by a ring of contrast-enhancement (CE) surrounding central necrosis. On the other hand, molecular glioblastomas are often, but not always, devoid of both CE and necrosis. Owing to their lower-grade histological features (grade 2 or 3), they primarily appear as FLAIR hyperintensities. This MRI appearance closely resembles that of a low-grade glioma (LGG), making the distinction between molGB and LGG on imaging very challenging (Figure 2).¹⁰ Characterizing them with quantitative morphometric analysis, by radiomics using machine learning or deep learning classifiers would help in differentiating molGB from LGG. In neuro-oncology, these methods have proved informative regarding the tumor, its prognosis, and its microenvironment.^11-17 To better characterize this subgroup of molecular glioblastoma, we studied their therapeutic response following radiation therapy combined with concurrent temozolomide according to the Stupp protocol, and compared their PFS and OS to those of patients with histGB treated with the same protocol, which is the standard care for glioblastoma. We also assessed the ability of traditional machine learning and deep learning models to differentiate molGB without CE from LGG.

Figure 2.

(A) Grade 2 astrocytoma, A₁ in T2-FLAIR sequence, A₂ T1 postgadolinium sequence. (B) Molecular glioblastoma, B₁ in T2-FLAIR sequence, B₂ in T1 postgadolinium sequence.

Methods

Survival data

Patient population

Between January 2018 and April 2022, we enrolled retrospectively patients with newly diagnosed glioblastoma (WHO 2021 classification), treated by radiotherapy at IUC-Oncopole Toulouse Cancer Center. Eligible patients must have received a concomitant fully radiotherapy course combined with oral temozolomide after surgical resection or biopsy. Surgery could be performed in any hospital; most procedures took place at the neurosurgery department of Purpan University Hospital in Toulouse. Patients who discontinued radiation therapy were treated elsewhere, or included in clinical trials were excluded. The design of the study was approved by the institutional review boards of our institution and was registered in Health Data Hub (N° F20230307142830).

Statistical analyses

Categorical variables were summarized as frequencies and percentages, while continuous variables were reported as medians and ranges. Group comparisons used the chi-squared or Fisher’s exact test for categorical variables and the Mann–Whitney test for continuous variables. Overall Survival (OS) was defined as the time from the date of diagnosis to the date of death from any cause. Progression-free survival was defined from diagnosis to progression or death from any cause. Patients still alive and progression-free were censored at last follow-up news. Survival rates were estimated using Kaplan–Meier analysis. Univariable and multivariable analyses were performed using the Logrank test and the Cox proportional hazards model, and hazard ratios (HR) were estimated with 95% CIs. All tests were two-sided, and P-values < .05 were considered significant. Analyses were conducted using Stata v16.

Imaging data and statistical modeling

Image preprocessing and radiomics

The FLAIR MRI preprocessing involved converting DICOM files to NifTI format, followed by skull-stripping using HD-BET,¹⁸ registration in the SRI24 anatomical template,¹⁹ and resampling for uniform isotropic resolution. This pipeline is publicly available via the BraTS Toolkit.²⁰ Segmentation masks were extracted using a deep learning model and then validated by a neuro-oncologist. The MONAI package²¹ was used to train a U-Net²² model from FLAIR MRI and segmentation masks from the BraTS database.²³ A single label, combining necrotic core, edematous/invaded tissue, and enhancing tumor, was used to enhance algorithm performance. Radiomic features from the pathological region of interest were extracted using the pyRadiomics,²⁴ computing 107 features to reduce spurious correlations, shown in the Table S1. These features included first-order statistics, texture analysis, and shape-based characteristics.

Deep learning modeling and pretraining

We used the ResNet10-3D architecture, effective in handling volume directly. To improve performance, we pretrained the encoder with FLAIR MRI from the BraTS dataset,²³ including both molecular glioblastoma and low-grade glioma, enabling the encoder to learn relevant features. The ResNet10-3D model was trained with a contrastive mechanism similar to SimCLR,²⁵ creating 2 distinct views for alignment. The projection head was an MLP with 512 units, followed by ReLU activation and a final 256-unit layer. The model was trained on full images and 64 × 64 × 64 volumes cropped around the segmentation mask.

Training and evaluation procedure

Radiomics

Experiments and modeling on radiomics were performed with the Scikit-learn,²⁶ for feature extraction, producing thousands of parameters per image. Three feature selection methods were applied: ANOVA F-Test, Mutual Information, and Recursive Feature Elimination. The top 10 features were selected to prevent overfitting and reduce computational cost. Three classifiers were tested to differentiate molGB and LGG: Random Forest (RF), Linear Support Vector Machine (L-SVM), and Logistic Regression (LR), with feature standardization for LR and L-SVM, and unchanged features for tree-based methods.

Deep learning

During training, pretrained encoder weights were frozen for whole MRI scans but not for tumor-centered scans. A classification head, consisting of 2 linear layers (256 units followed by ReLU activation), was added to distinguish between molGB and LGG. The Adam optimizer was used with a learning rate of 1e−4, and training lasted 15-30 epochs with a batch size of 6. Data augmentations, including flipping, rotations, and dropout, were applied to enhance robustness. The dataset was split into 3 folds for cross-validation, repeated 100 times with different fold compositions. Each example was used 200 times in training and 100 times in testing. Model performance was averaged over 300 runs to obtain trends without hyperparameter optimization.

Results

Survival data

Baseline characteristics

After excluding patients who discontinued radiotherapy or where in clinical trials, 132 patients treated by radio-chemotherapy were included: 103 with histGB (78%) and 29 with molGB (22%). Baseline characteristics (Tables 1 and 2) showed no significant differences in sex, performance status (PS), tumor location, or MGMT promoter methylation. Patients with molGB were slightly younger (median 65 vs 68 years, P = .049).

Table 1.

Patient characteristics.

	Total (N = 132)	histGB (N = 103)	molGB (N = 29)	P-value
Sex, n (%)				.200
Male	87 (65.9%)	65 (63.1%)	22 (75.9%)
Female	45 (34.1%)	38 (36.9%)	7 (24.1%)
Age at diagnosis, y				.048
Median	67.5	68.0	65.0
(Range)	(29.0-86.0)	(36.0-86.0)	(29.0-86.0)
Age, y				.024
<70	86 (65.2%)	62 (60.2%)	24 (82.8%)
>= 70	46 (34.8%)	41 (39.8%)	5 (17.2%)
Preoperative PS				.365
0-1	106 (80.3%)	81 (78.6%)	25 (86.2%)
2-3	26 (19.7%)	22 (21.4%)	4 (13.8%)
Localization, n				.084
Frontal	36 (27.3%)	31 (30.1%)	5 (17.2%)
Parietal	29 (22.0%)	23 (22.3%)	6 (20.7%)
Temporal	54 (40.9%)	41 (39.8%)	13 (44.8%)
Occipital	4 (3.0%)	4 (3.9%)	0 (0.0%)
Thalamus	2 (1.5%)	1 (1.0%)	1 (3.4%)
Brainstem	2 (1.5%)	0 (0.0%)	2 (6.9%)
Insula	4 (3.0%)	2 (1.9%)	2 (6.9%)
Cerebellar	1 (0.8%)	1 (1.0%)	0 (0.0%)
Hemisphere, n				.892
Right	62 (47.7%)	48 (46.6%)	14 (51.9%)
Left	66 (50.8%)	53 (51.5%)	13 (48.1%)
Both	2 (1.5%)	2 (1.9%)	0 (0.0%)
Missing	2	0	2
MGMT promoter methylation status, n				.472
Unmethylated	66 (54.5%)	55 (56.1%)	11 (47.8%)
Methylated	55 (45.5%)	43 (43.9%)	12 (52.2%)
Missing	11	5	6
MIB1 (%)				<.0001
Median	22.5	25.0	15.0
(Range)	(1.0-90.0)	(10.0-90.0)	(1.0-40.0)
Missing	10	5	5
Contrast enhancement on diagnostic MRI, n
No	19 (14.4%)	1 (1.0%)	18 (62.1%)
Yes	113 (85.6%)	102 (99.0%)	11 (37.9%)

Bold corresponds to statistically significant difference.

Table 2.

Treatment characteristics.

	Total (N = 132)	histGB (N = 103)	molGB (N = 29)	P-value
Surgical modality, n				.004
Biopsy	17 (12.9%)	8 (7.8%)	9 (31.0%)
Partial resection	45 (34.1%)	35 (34.0%)	10 (34.5%)
Total resection	70 (53.0%)	60 (58.3%)	10 (34.5%)
Time between surgery-radiotherapy, months (range)				.940
Median	1.4 (0.9-2.6)	1.4 (0.9-2.1)	1.4 (0.9-2.6)
Total dose received, n
60 Gy	119 (90.2%)	94 (91.3%)	25 (86.2%)
40 Gy	13 (9.8%)	9 (8.7%)	4 (13.8%)
Full concomitant TMZ, n	131 (99.2%)
Adjuvant TMZ, n cycles (range)				.516
Median	5 (0.0-15.0)	5 (0.0-15.0)	6 (0.0-12.0)
MRI evaluation at 1 month, n				.085
Stability	57 (43.8%)	40 (39.6%)	17 (58.6%)
Increased contrast	54 (41.5%)	47 (46.5%)	7 (24.1%)
Partial Regression	19 (14.6%)	14 (13.9%)	5 (17.2%)
Missing	2	2	0
Pseudo-Progression, n				1.000
Yes	22 (17.9%)	18 (18.6%)	4 (15.4%)
No	101 (82.1%)	79 (81.4%)	22 (84.6%)
Missing	9	6	3
2nd-line treatment for patients progressing, n	105	84	21	.810
None	19 (18.8%)	15 (18.5%)	4 (20.0%)
Surgery	29 (28.7%)	25 (30.9%)	4 (20.0%)
Re-irradiation	13 (12.9%)	10 (12.3%)	3 (15.0%)
Systemica	40 (39.6%)	31 (38.3%)	9 (45.0%)
Missing	4	3	1
Median follow-up, months (range)	24.6 (21.6-28.8)	25.2 (21.6-28.8)	24.6 (16.8-31.8)

^aChemotherapy, targeted therapy, or combination of the two.

Abbreviation: TMZ, temozolomide.

Glioblastoma typically exhibits irregular, peripheral CE on gadolinium-enhanced MRI, with a ring-like pattern, reflecting blood–brain barrier disruption, with necrotic central area. In our cohort, 99% of histGB exhibited these characteristics, whereas only 38% of patients in the molGB group had tumors with such CE. This means that 62% of the tumors observed in the molGB group appeared only as a FLAIR hyperintensity on MRI.

The management of glioblastoma or any other high-grade tumor is primarily based on surgery. This serves both an etiological purpose by confirming the histological diagnosis, and, when possible, a therapeutic purpose, including biopsy, partial resection, or subtotal resection. In the case of glioblastoma with CE, subtotal resection is generally considered to have been achieved when all the contrast-enhancing tumor tissue has been removed. When the tumor does not enhance with contrast, complete resection refers to the removal of the entire FLAIR hyperintensity, in accordance with the analogy with low-grade tumors.

Around one-third of patients with molGB underwent a biopsy, one-third had a partial resection and one-third a subtotal resection. Regarding patients with histGB, more than 50% had a subtotal resection and only 7.8% were biopsied (P = .004). However, it should be noted that the definition of subtotal resection differs between the 2 groups.

Time to radiotherapy after surgery (median 1.4 months) and total radiotherapy dose (mainly 60 Gy) were similar in both groups, with fewer than 15% receiving hypofractionated regimens. Most patients completed temozolomide without interruption, with a median of 6 adjuvant cycles for molGB and 5 for histGB (P = .517). No significant differences were observed in second-line treatments after progression.

Molecular glioblastomas are defined as grade 2 or 3 astrocytomas that exhibit specific molecular characteristics. These molecular markers include TERT promoter mutations or EGFR amplification, or the combined gain of the entire chromosome 7 and loss of the entire chromosome 10 [+7/−10]. The presence of any of these, or their combination, in IDH-wild-type diffuse astrocytomas allows classification as glioblastoma, IDH-wild-type CNS WHO grade 4, even if they appear histologically lower-grade. By examining the distribution of these molecular markers in the molGB group, we observed that 27.5% had EGFR amplification, 20.6% had a chr +7/−10, and over 95% presented a TERT promoter mutation. Regarding the distribution of histological grade among molGB, 38% were grade 2 and 62% were grade 3 based on their morphological features. Regarding the correlation with their MRI characteristic at diagnosis, about two-thirds of molGB showed no CE. Among these, 59% were grade 3 and 41% were grade 2 according to their morphological features.

Survival outcome

Median follow-up was 24.6 months (%95 CI, 21.6-28.8). Median OS was 18.6 months (95% CI, 16.3-20.1) with no significant difference between molGB was 20.6 months and histGB 18.4 months (HR 0.70, 95% CI, 0.40-1.22, P = .2). However, in univariable analysis, molGB without CE was significantly associated with longer survival than histGB (HR 0.45, 95% CI, 0.21-0.99) with respectively 31.2 months and 18 months of median OS, while molGB with CE showed no significant difference (HR 1.31, 95% CI, 0.63-2.74). After adjustment on clinical factors, the results were similar (HR molGB without CE vs histGB: 0.36, 95% CI, 0.13-1.04; HR molGB with CE vs histGB: 1.75, 95% CI, 0.72-4.27). Patients with PS 2 and 3 were associated with worse prognosis in univariable analysis (HR 1.91, 95% CI, 1.15-3.16, P = .011). As expected, patients with methylated MGMT glioblastoma had a significantly better outcome (P = .007) with a median OS of 20.6 months (95% CI, 17.6-NR) vs 18.6 months (95% CI, 14.8-19.7) for those without. These differences were still statistically significant in multivariable analysis. Survival was not significantly affected by the type of surgery (P = .453).

The survival curves are shown in Figure 3. Median PFS was 9.5 months (95% CI, 8.2-10.3): 9.3 months (95% CI, 7.5-10.3) for histGB, and 10.1 months (95% CI, 9.4-12.5) for molGB, with an HR of 0.73 (95% CI, 0.45-1.17) P = .183. Considering the subgroup of patients with tumors lacking contrast, PFS was not different (HR molGB without CE versus histGB: 0.66, 95% CI, 0.37-1.19) and they had a median PFS of 11.3 months (95% CI, 8.8-13.8) vs 9.3 months (95% CI, 7.5-10.3). In neither univariable nor multivariable analysis were PS or surgery modality associated with PFS.

Figure 3.

(A) OS and PFS between molGB and histGB, in univariable analysis. (B) OS and PFS between molGB without CE, histGB and molGB with CE, in univariable analysis.

Radiomics and deep learning findings

We subsequently assessed the capacity of AI and statistical modeling to differentiate molGB without CE from LGG in 41 patients (21 with molGB and 20 with IDH-mutated grade 2 astrocytomas), using noncontrast-enhancing MRI to compare T2-FLAIR hyperintensity features. We used traditional machine learning on radiomics features and deep learning on either whole MRI scan or tumor regions only. We developed 12 radiomics models utilizing 3 machine learning algorithms and 3 feature selection methods, including an approach without feature selection. The diagnostic performance of the top traditional and deep learning models is reported in Table 3.

Table 3.

Diagnostic performance of traditional machine learning and deep learning models. AUC: area under receiver operating characteristic curve.

	AUC	Precision	Recall	Accuracy
Radiomics-based models
LR with F-score selection	0.66 ± 0.16	0.64 ± 0.14	0.71 ± 0.18	0.63 ± 0.13
L-SVM with RFE selection	0.64 ± 0.14	0.63 ± 0.13	0.64 ± 0.19	0.61 ± 0.12
LR with MI selection	0.63 ± 0.15	0.60 ± 0.14	0.67 ± 0.19	0.59 ± 0.13
RandomForest with all features	0.67 ± 0.13	0.61 ± 0.14	0.60 ± 0.19	0.60 ± 0.13
Deep learning-based models
Tumor ResNet10-3D pretrained	0.75 ± 0.12	0.65 ± 0.18	0.71 ± 0.29	0.64 ± 0.12
ResNet10-3D pretrained	0.85 ± 0.07	0.71 ± 0.10	0.88 ± 0.13	0.74 ± 0.08

In radiomics-based analyses, the RF model without feature selection achieved the highest AUC reaching 0.67 ± 0.13. Despite this marginal advantage, the overall performance of the models remained similar, with an AUC of around 0.65, regardless of feature selection. The 4 radiomic models produced similar results, with moderate performance in terms of precision (between 0.60 and 0.64), recall (between 0.60 and 0.71) and accuracy (between 0.59 and 0.63), although the aim was not to compare the models with each other.

In deep learning assessments, focusing on tumor regions slightly improved results over radiomics-based models, particularly in AUC (0.75 ± 0.12), though recall and precision showed broader confidence intervals (0.71 ± 0.29 and 0.65 ± 0.18, respectively). Finally, the deep learning model trained on entire MRI scans without isolating hyperintense areas outperformed traditional methods in diagnostic accuracy, reaching an average AUC of 0.85 ± 0.07. The deep learning model also outperformed radiomics models particularly in its recall of 0.88 ± 0.13. This underlines its sensitivity and thus its ability to identify true glioblastomas, thereby minimizing the risk of false negatives during prediction.

Discussion

This retrospective study evaluated the clinical outcome of glioblastomas lacking necrosis and microvascular proliferation, aka. “molecular glioblastomas” (molGB), in comparison with histological grade 4 glioblastomas (histGB). Furthermore, we investigated the ability of radiomics and deep learning models to differentiate them from LGG. To our knowledge, this is the first study evaluating the survival of patients with molGB receiving the standard treatment for glioblastoma, ie, chemoradiation according to the Stupp protocol.

The cIMPACT-NOW committee introduced a novel subtype of glioma known as diffuse astrocytic glioma, IDH1/2 wild-type, exhibiting molecular characteristics similar to glioblastoma WHO grade 4 (4). Despite harboring histological features that would previously be rated as lower grade, the presence of pTERT mutation status or EGFR amplification or combined gain of chromosome 7 and loss of chromosome 10 (+7/−10) is now sufficient to assign the highest WHO grade, and thus the diagnosis of IDH-wild-type glioblastoma, despite the absence of tumor necrosis or vascular proliferation.^27,28 At present, these “molecular glioblastomas” are no longer differentiated from histologically defined glioblastomas in the 2021 WHO classification. They are fully comingled as a single entity termed “glioblastoma IDH wildtype, WHO grade 4” and they all receive the same standard treatment.⁹ The survival outcome of these glioblastomas lacking necrosis or vascular proliferation remains controversial.²⁹

All the patients in our 2 cohorts received the same standard treatment, ie, concomitant chemoradiation according to the Stupp protocol, followed by Temozolomide maintenance, with no significant difference in treatment modalities. The median age at diagnosis was slightly higher in patients with histGB (68 years vs 65 years, P = .049). However, the proportion receiving hypofractionated treatment was similar. Median OS in patients with molGB was 20.6 months (95% CI, 15.6-31.2) and 18.4 months (95% CI, 14.9-19.9) in those with other glioblastomas (HR 0.70, 95% CI, 0.40-1.22, P = .2). Moreover, PFS was not significantly different (P = .18), which is consistent with the current WHO classification grouping them as a single entity.

By definition, molGB have fewer nuclear atypia, a lower proliferation index and are characterized by the absence of necrosis or microvascular proliferation. Therefore, the MIB1 proliferative index was significantly higher in the histGB group (P < .0001), given their prognostic indicator for high‑grade glioma, with a strong correlation with outcome and survival.²⁹ Despite these histological differences, however, we did not find any survival advantage in favor of the molGB group.

The lower rate of subtotal resection in patients with molGB (34.5% vs 58.3% in patients with histGB) reflects challenges in resecting infiltrative, multilobar lesions. However, the definitions of what constitutes subtotal resection differ. In histGB, it includes resection of CE areas, while in molGB without CE, it concerns removal of FLAIR extension. Caution is thus required when comparing outcomes. The variation in resection rates might explain the limited outcomes of patients with molGB, given the importance of gross tumor resection as a prognostic factor. Currently, there are no guidelines for resecting nonenhancing glioblastomas. Nonetheless, our multivariable analysis that included surgical techniques shows that survival rates remain comparable between histGB and molGB. Tesileanu et al. reported similar survival rates between molGB and histGB (HR 1.27, 95% CI, 0.85-1.88, P = .242). However, only 42% of patients with molGB received appropriate chemoradiation, potentially biasing survival outcomes.³⁰ Wijnenga et al. also found no survival advantage for molGB, but the study lacked details on treatment modalities.³¹

Our analysis revealed a median PFS in patients with molGB without CE of 11.3 months (95% CI, 8.8-13.8) vs 9.3 months (95% CI, 7.5-10.3) in patients with histGB, with an HR of 0.66 (95% CI, 0.37-1.19) and a median OS of 31.2 months (95% CI 19.6-NR) versus 18.4 months (95% CI, 14.9-19.9) respectively, with an HR of 0.45 (95% CI 0.21-0.99). This suggests a favorable outcome when patients are treated with the Stupp protocol. Moreover, since almost one-third of the patients with molGB without CE were biopsied, our results suggest that this subgroup is more sensitive to radio chemotherapy than the other patients. These tumors are also naturally more indolent, which may also partly explain this finding. This result is consistent with previous studies showing that the proportion of CE on MRI is associated with survival in glioblastoma.³² Pope and al. demonstrated that the absence of enhancing tumor, edema, and satellite or multifocal lesions, is correlated with a doubling of the median survival.¹⁰

Among patients with noncontrast molecular glioblastoma, 59% were classified as histological grade 3 based on their higher proliferative activity compared to grade 2, so their prognosis was worse. Owing to the limited sample size, we were unable to compare the survival rates between histological grade 2 and grade 3 molecular glioblastomas. However, it would be interesting to determine whether histological grade within the molecular cohort directly influences their prognosis. In Berzero’s study, IDH wild-type grade 2 gliomas with molecular glioblastoma features had an OS of 42.2 months, 28% of them having received the Stupp protocol. Therefore, grade 3 histology is aligned with glioblastoma outcomes.³³ Conversely, although most of our patients were grade 3, results suggest a longer OS in patients with nonenhancing tumors, regardless of histological grade.

Despite the paucity of literature on molGB and its absence of distinction in the WHO classification, documenting these tumors remains crucial, ideally in larger cohorts. Indeed, most recent and ongoing trials do not differentiate molGB, potentially biasing survival rates and treatment responses. The optimal chemoradiation regimen, surgical approach, imaging features, and effectiveness of Tumor-Treating Fields for molGB without CE remain unclear. Glioblastomas typically demonstrate contrast-enhancement on T1-gadolinium MRI, with irregular borders, necrosis, and T2-FLAIR hyperintensity. Although nonenhancing lesions are usually evocative of LGG, most molGB do not enhance, which may delay biopsy or resection, diagnosis, and treatment initiation.³⁴

Artificial intelligence (AI), including deep learning and machine learning, addresses human limitations in imaging analysis. Artificial intelligence plays a growing role in neuro-oncology, by enabling automated quantitative assessment of complex imaging data. Artificial intelligence have demonstrated its ability to distinguish glioblastoma from gliosarcoma, primary central nervous system lymphoma,³⁵ brain metastases,³³ and radiation necrosis.³⁶ Some studies have also successfully used machine learning to predict WHO grade.³⁷ However, these models often compared tumors with obvious different imaging characteristics, such as necrosis or contrast-enhancement, which could explain their high accuracy.

The ability of these models to distinguish tumors with similar imaging characteristics that are almost indistinguishable to the human eye offers much promise. Our study assessed their ability to differentiate molGB without contrast-enhancement from Low-Grade Glioma, on T2-FLAIR sequences. All patients included had the same imaging characteristics such as lack of contrast and flair hypersignal.

Despite a small dataset, statistical modeling showed promise with a deep learning model utilizing broader MRI data outperforming tumor-focused approaches (AUC: 0.85). While the models were not optimized for peak performance, they highlight AI’s potential for noninvasive tumor analysis, aiding diagnosis in nonoperable patients and reducing reliance on observer expertise.

The main limitations of our study include its retrospective design, small sample size, and single-center nature, which, while ensuring consistent treatment and follow-up, may lead to sampling bias and overfitting. Challenges in machine learning, particularly in repeatability and reproducibility, further limit the impact of our findings. The lack of consensus on reliable radiomics features and the influence of image acquisition and segmentation on model performance add complexity.^38-40 Despite these barriers, radiomics and AI remain promising for brain tumor diagnosis and prognosis. While not yet applicable to clinical practice, our results highlight the need for larger, multicenter studies.

Conclusion

The survival of patients with molGB and histologic grade 4 glioblastomas was similar in this cohort. However, when treated with the Stupp protocol, patients with noncontrast molGB seem to achieve better outcomes compared to those with classic glioblastomas, suggesting that stratification based on the absence of contrast enhancement should be considered in randomized clinical trials. Further prospective studies are needed to better understand this molGB. In addition, feature-based and deep learning-based radiomics hold considerable potential for providing important diagnostic information to answer many highly relevant clinical questions in brain tumor patients such as those with molGB.

Author contributions

Conception/design: E.C.J.M.; Provision of study material or patients: M.R., D.L., A.S., and F.R.; Collection and/or assembly of data: C.Z., L.R., A.S, and S.K.; Data analysis and interpretation: C.Z., L.R., S.K., A.C., A.B., and E.C.J.M.; Manuscript writing: C.Z., L.R., and E.C.J.M.; Final approval of manuscript: all authors.

Funding

None declared.

Conflicts of interest

E.C.J.M. served as an expert board member for Novocure and received lecture fees from Accuray and Novocure, travel expenses from Novocure, and research grants from Astra Zeneca, Novocure, Bayer, and Incyte. She also received research grants from the ARC Foundation. All the other authors have no conflict of interest. No disclosures are reported by the other authors.

Data availability

All data described in this study are freely available for academic use and can be obtained on request to the corresponding author by email.