Sage Journals: Discover world-class research

Abstract

French

Purpose: To evaluate the value of intra- and peritumoural deep learning (DL) features based on multi-parametric magnetic resonance imaging (MRI) for identifying telomerase reverse transcriptase (TERT) promoter mutation in glioblastoma (GBM). Methods: In this study, we included 229 patients with GBM who underwent preoperative MRI in two hospitals between November 2016 and September 2022. We used four 2D Convolutional Neural Networks (GoogLeNet, DenseNet121, VGG16, and MobileNetV3-Large) to extract intra- and peritumoural DL features. The Mann–Whitney U test, Pearson correlation analysis, least absolute shrinkage and selection operator, and logistic regression analysis were used for feature selection and construction of DL radiomics (DLR) signatures in different regions. These multi-parametric and multi-region signatures were combined to identify TERT promoter mutation. The area under the receiver operating characteristic curve (AUC) was used to evaluate the effects of the signatures. Results: The signatures based on the DL features from the peritumoural regions with expansion distances of 2 mm, 8 mm, and 10 mm using the GoogLeNet architecture correlated with the optimal AUC values (test set: .823, .753, and .768) in the T2-weighted, T1-weighted contrast-enhanced, and T1-weighted images. Using the stacking fusion method, DLR with multi-parameter and multi-region fusion achieved the best discrimination with AUC values of .948 and .902 in the training and test sets, respectively. Conclusions: The radiomics model based on the fusion of multi-parameter MRI intra- and peritumoural DLR signatures may help to identify TERT promoter mutation in patients with GBM.

Keywords

glioblastoma telomerase reverse transcriptase promoter magnetic resonance imaging deep learning radiomics

Introduction

Glioblastoma (GBM) is the most common and most aggressive primary tumour of the central nervous system (CNS). Its incidence varies by country and population, with an overall annual incidence of 3.26 cases per 100,000 inhabitants.¹ Despite advances in the management of GBM, patient prognosis remains poor with a mean overall survival of 16-17 months.² According to the fifth edition of the WHO Classification of Tumors of the CNS (WHO CNS5), molecular markers are crucial for determining accurate brain tumour diagnosis and predicting patient outcome.³ The criteria for GBM diagnosis include adult-type diffuse gliomas with microvascular proliferation, necrosis, and specific molecular features, such as telomerase reverse transcriptase (TERT) promoter mutation, epidermal growth factor receptor amplification, chromosome 7 gain, and chromosome 10 loss.⁴ Therefore, TERT promoter mutation status plays an important role in GBM diagnosis.⁵

Although complete resection of GBM contrast-enhancing lesions is associated with prolonged survival,⁶ the tumour often recurs due to the presence of tumour cells beyond the contrast-enhancing lesions.⁷ GBM with TERT promoter mutation is more aggressive and has greater peritumoural tumour cell infiltration; thus, maximum surgical resection of contrast-enhancing and non-contrast-enhancing lesions should be performed when safe and feasible.^8,9 Therefore, a preoperative identification of TERT promoter mutation may play a key role in formulating an appropriate surgical plan.

With advances in machine learning algorithms and computer technology, computer-aided diagnosis has become an important area of medical imaging research. Previous studies on radiomics have revealed that peritumoural features in medical images have good diagnostic ability.^10,11 Models for predicting grade in gliomas incorporate peritumoural radiomics features with satisfactory results.¹² Peritumoral regions contain valuable information that is closely linked with tumour characteristics.¹³ Furthermore, studies have shown that using a machine learning model based on deep learning (DL) features from multi-parametric magnetic resonance imaging (MRI) can achieve good classification results.¹⁴ This indicates that DL features are associated with high selectivity, stability, and robustness in distinguishing normal from abnormal patterns.¹⁵ However, whether the intra- and peritumoural DL features can be used for non-invasive evaluation of TERT promoter mutation in GBM patients is currently underexplored. We hypothesized that DL features of the intra- and peritumoural regions would help identify TERT promoter mutation in patients with GBM.

The aim of this study was to develop and validate a radiomics model that incorporated intra- and peritumoural DL features of multi-parametric MRI to preoperatively identify TERT promoter mutation in patients with GBM.

Methods

Patients

This study was approved by our Institutional Ethics Committee, and the requirement for informed consent was waived due to the retrospective nature of this multicenter study. Data from patients with histologically confirmed isocitrate dehydrogenase (IDH)-wild-type GBM (WHO grade 4) between November 2016 and September 2022 were included in this study. The data were obtained from the relevant medical institution I and II databases. Figure 1 and Supplementary Method S1 present the patient recruitment process, inclusion, and exclusion criteria. In total, 229 GBM patients in this study were included. The patients were divided into a training set (115 patients [72 men and 43 women; mean age, 54.3 ± 13.4 (21-96) years] from medical institution I) and an external test set (114 patients [72 men and 42 women; mean age, 54.5 ± 12.4 (23-86) years] from medical institution II). Baseline clinical characteristics included age, sex, tumour location, number of recurrences, and mean postoperative progression-free survival time. Mutation hotspots in the IDH1, IDH2, and TERT promoters (C228 T and C250 T) were detected by Sanger sequencing.

Figure 1.

Data collection flow chartThe inclusion and exclusion criteria for selecting the study populations are shown for the training and test sets. GBM, Glioblastoma; TERT, Telomerase reverse transcriptase.

MRI protocol and preprocessing

Figure 2 summarizes the workflow employed. GBM magnetic resonance (MR) images [axial T1-weighted contrast-enhanced image (T1CE), T1-weighted image (T1WI), and T2-weighted image (T2WI)] were retrieved from a Picture Archiving and Communication System (PACS) for further evaluation. The following MRI acquisition protocol was used (Supplementary Table S2).

Figure 2.

Overall study flowchart. ROI, Region of interest; CNN, Convolutional neural network; DL, Deep learning; DLR, Deep learning radiomics; FDLR, Fusion deep learning radiomics; LR, logistic regression.

To pre-process the image data, we first performed N4 bias field correction on the collected MR images using the SimpleITK package (NumFOCUS, Austin, TX). The T1WI and T1CE were then registered as T2WI using FLIRT (action options: Affine, 12 parameter model, and trilinear interpolation method) in the FMRIB Software Library (Analysis Group, FMRIB, Oxford, UK). Next, the images were re-sampled to a uniform voxel size of 1 × 1 × 1 mm³ using the nearest-neighbor interpolation method of the “SimpleITK” package. Finally, the image signal intensity was normalized using the “PyRadiomics” package.¹⁶

Tumour Segmentation

Reader 1, with 3 years of experience in neurological MRI interpretation, first delineated the tumour’s region of interest (ROI) based on the MRI images of 229 patients using the ITK-SNAP software (version 3.6.0; Penn Image Computing and Science Laboratory, Philadelphia, PA). The outline was defined as the tumour core (tumour enhancement, necrosis, and cystic region), which included all tumour structures, as shown in the T1CE and T2WI. For tumours lacking an enhancement area, the tumour core with an intermediate signal on T1WI and T2WI was delineated. Reader 2, with 7 years of experience in neurological MRI interpretation, validated the ROIs for all images. Reader 3, with 30 years of experience in neurological MRI interpretation, provided the validation in cases of disagreements between Reader 1 and Reader 2.

After determining the core of the tumour and to further evaluate the predictive ability of the tumour core's surrounding area, the originally masked area was expanded in segments at 2 mm intervals to five radial distances outside the area of the tumour core (maximum expansion distance was 10 mm) by Onekey AI platform (OnekeyAI, Shijiazhuang, Hebei). Figure 3 shows the mask expansion process, and the final mask shows the intra- and peritumoural regions. Previous studies have shown that the surrounding 10 mm peritumoural area outside of the tumour core can provide GBM-related diagnostic information.¹⁷ Therefore, selecting the parenchyma 10 mm around the tumour at the imaging level is reasonable for tumour behaviour stratification.^18,19

Figure 3.

MRI masks. An example of dilated MRI masks for different radial dilation distances outside the original mask is shown. The red area represents the segmented tumour core area and the coloured rings indicate the peritumoral area, with each ring representing a 2 mm wide dilation. MRI, magnetic resonance imaging.

Extraction and Selection of DL Features

DL features were extracted using four 2D convolutional neural networks (CNNs), including GoogLeNet, DenseNet121, VGG16, and MobileNetV3-Large. The code for each architecture was derived from previous research.^20

-23 After pre-training all the 2D CNNs using ImageNet (Stanford Vision Lab, Stanford University, Palo Alto, CA), the largest horizontal axial ROI images of the intra- and peritumoural regions in each sequence image were used for feature extraction. DenseNet121 extracted the features of the features.norn5 layer, whereas the rest of the CNN architecture used an average pooling layer to extract the DL features. Finally, principal component analysis was used for deep feature compression, and 200 DL features were output from the intra- and peritumoural regions of the T1CE, T1WI, and T2WI.

Feature selection first used Z-score normalization to standardize the features, which returned all data to a distribution with a mean of zero and a variance of one. Feature selection was performed on the training set as follows: (1) the Mann–Whitney U test was used to retain features with a P-value <.05; (2) The Pearson correlation analysis was performed, and the correlation coefficients among features were calculated to evaluate their multi-collinearity. If there was any coefficient value ≥.9 for a pair of features, then only the feature with a better diagnostic performance was retained; (3) the key features were selected using the least absolute shrinkage and selection operator (LASSO). The optimal regulation weight λ was found using 10-fold cross-validation with minimum criteria. The Python “scikit-learn” package (Python Software Foundation, Fredericksburg, VA) was used for LASSO regression modelling.

Differentiation of TERT Promoter Status With Different Signatures

We aimed to assess the discriminative ability of intra- and peritumoural regions on T1CE, T1WI, and T2WI to identify the presence of TERT promoter mutation in patients with GBM. The key features of the selected multi-parametric MRI and multi-ROI regions were constructed with corresponding deep learning radiomics (DLR) signatures. The signature of the best prediction result was then selected for different expansion distances for each MRI sequence to construct the fusion DL radiomics (FDLR) signature. We employed two multi-parametric fusion strategies: feature concatenation and stacking fusion. Feature concatenation involved concatenating the DL features extracted from different ROIs of each MRI sequence to obtain fused multi-parametric features for modelling. Stacking fusion combined the DLR signatures on different ROIs of each MRI sequence using another machine learning algorithm.

To validate the performance of the FDLR signature with multi-parametric MRI, only signatures with either one or two parameters were used to determine the presence of TERT promoter mutation in GBM patients.

Statistical Analysis

All statistical analyses were performed using Python software (version 3.7.14, Python Software Foundation, Fredericksburg, VA). The receiver operating characteristic (ROC) curve and area under the curve (AUC) were used to evaluate the signatures with different sequences, ROIs, and fusion strategies. Other evaluation metrics included accuracy, sensitivity, specificity, positive predictive value, negative predictive value, precision, recall, and F1-score. The Delong test was used to calculate the statistical significance between two ROC curves.

The clinical characteristics of patients in the different groups were compared using t-tests for continuous variables and chi-squared or Fisher’s exact tests for categorical variables, as appropriate. Statistical significance was set at a P-value <.05.

Results

Patient Characteristics

Table 1 summarizes the baseline clinical characteristics of the enrolled patients with GBM. In the training and external test sets, the TERT promoter mutation rates were 54.8% (63/115) and 64.0% (73/114), respectively. The average initial age at onset in the TERT promoter mutant group was older than that in the wild-type group in both the training (50.7 ± 13.3 vs 57.3 ± 12.8) and external test sets (50.0 ± 12.6 vs 57.1 ± 11.7). The two groups had no other significant baseline clinical differences (P > .05).

Table 1.

Clinical Characteristics of Glioblastoma Patients.

Characters	Train (ALL)	Train (W)	Train (M)	P-value	Test (ALL)	Test (W)	Test (M)	P-value
TERT promoter	115	52	63		114	41	73
Age	54.3 ± 13.4	50.7 ± 13.3	57.3 ± 12.8	.008*	54.5 ± 12.4	50.0 ± 12.6	57.1 ± 11.7	.003*
Sex				.551				.212
Female	43	21	22		42	12	30
Male	72	31	41		72	29	43
Location				.590				.370
Frontal lobe	31	14	17		19	6	13
Occipital lobe	2	0	2		3	1	2
Parietal lobe	17	10	7		14	6	8
Temporal lobe	20	9	11		42	11	31
Insular lobe	2	1	1		3	0	3
Basal nuclei or thalamus	4	3	1		9	7	2
Brainstem or cerebellum	3	1	2		5	5	0
Across multilobe	36	14	22		19	5	14
Number of recurrences	41	17	24		43	13	30
Mean progression-free survival time	7.9 ± 14.0	7.6 ± 12.6	8.1 ± 15.0	.814	7.5 ± 6.4	7.3 ± 8.1	7.6 ± 5.6	.884

*P-value <.05 indicates statistical significance.

M, Telomerase reverse transcriptase promoter mutation; W, Telomerase reverse transcriptase promoter wild-type.

Evaluation of Intra- and Peritumoural Regions

Table 2 lists the effects of DLR signatures constructed by the logistic regression model on the discrimination of TERT promoter mutation based on GoogLeNet, DenseNet121, VGG16, and MobileNetV3-Large for extracting DL features in the intra- and peritumoural regions in T1CE, T1WI, and T2WI. The bold font indicates the DLR signature with the best predictive effect on different ROI regions of each 2D CNN in the three MRI sequences, high AUC values, and low overfitting. In T2WI, T1CE, and T1WI, the GoogLeNet DLR signatures had the best resolution in the intra- and peritumoural regions with dilation distances of 2 mm, 8 mm, and 10 mm and test set AUCs of .823, .753, and .768, respectively. The DenseNet121 DLR signatures showed the best classification effect in the intra- and peritumoural regions, with expansion distances of 8 mm, 8 mm, and 10 mm and AUCs of .762, .788, and .821 in T1CE, T1WI, and T2WI, respectively. The VGG16 DLR signatures showed the best discriminative ability in the intra- and peritumoural regions, with dilation distances of 2 mm, 6 mm, and 8 mm and AUCs of .798, .792, and .776 in T1CE, T2WI, and T1WI, respectively. The MobileNetV3-Large DLR signatures showed excellent evaluation ability in the intra- and peritumoural areas, with expansion distances of 0 mm, 4 mm, and 10 mm and AUCs of .793, .751, and .830 in T1CE, T2WI, and T1WI, respectively.

Table 2.

TERT Promoter Mutation Prediction Performance of Intra- and Peritumoural Regions in T1CE, T1WI, and T2WI.

Sequence	Dilation Distance (mm)	GoogLeNet	DenseNet121	VGG16	MobileNetV3-Large	SET
Sequence	Dilation Distance (mm)	AUC	AUC	AUC	AUC	SET
T1CE	0	.781	.784	.799	.815	Train
	0	.712	.721	.745	.793	Test
	2	.807	.808	.781	.818	Train
	2	.707	.745	.798	.672	Test
	4	.735	.792	.797	.861	Train
	4	.732	.666	.733	.721	Test
	6	.871	.895	.777	.777	Train
	6	.660	.710	.723	.700	Test
	8	.835	.881	.813	.837	Train
	8	.753	.762	.779	.686	Test
	10	.803	.856	.719	.881	Train
	10	.738	.740	.705	.664	Test
T1WI	0	.825	.647	.834	.854	Train
	0	.766	.630	.700	.754	Test
	2	.815	.831	.765	.883	Train
	2	.696	.695	.768	.679	Test
	4	.850	.777	.748	.811	Train
	4	.744	.650	.717	.721	Test
	6	.751	.858	.864	.834	Train
	6	.676	.716	.747	.748	Test
	8	.751	.778	.791	.801	Train
	8	.674	.788	.776	.715	Test
	10	.796	.744	.792	.898	Train
	10	.768	.768	.746	.830	Test
T2WI	0	.903	.796	.773	.831	Train
	0	.764	.742	.643	.701	Test
	2	.864	.776	.811	.825	Train
	2	.823	.645	.733	.698	Test
	4	.811	.799	.834	.842	Train
	4	.759	.695	.691	.751	Test
	6	.815	.753	.866	.685	Train
	6	.712	.759	.792	.751	Test
	8	.826	.920	.811	.799	Train
	8	.772	.800	.734	.730	Test
	10	.797	.812	.786	.909	Train
	10	.761	.821	.776	.694	Test

AUC, Area under receiver operating characteristic curve.

Evaluation of Different Multi-Parametric Fusion Strategies

In the training and external test sets, the FDLR signatures using the feature concatenation and stacking fusion strategy on all evaluation indicators showed the ability to discriminate between TERT promoter mutation and wild-type (Table 3 and Figure 4). The feature concatenation strategy used the logistic regression for model building. The stacking fusion strategy used the logistic regression for signature fusion. Among them, the FDLR signature constructed by GoogLeNet DLR signatures using the stacking fusion strategy achieved the best evaluation performance: the AUCs for the training and external test sets were .948 (95% CI, .911-.985) and .902 (95% CI, .843-.961), respectively. The confusion matrix for each model is shown in Supplementary Figure S3. There were no significant differences in the results when the two fusion strategies were used in the same 2D CNN architecture (P < .05) (Supplementary Table S4). However, the stacking fusion strategy minimized the relatively large differences caused by overfitting between the training and external test sets in the feature concatenation fusion method.

Table 3.

Differentiated TERT Promoter Mutation Status Performance of Different Fusion Strategies in the Training and Test Sets.

Model name	ACC	AUC	95% CI	SEN	SPE	PPV	NPV	PRE	Recall	F1	SET
GoogLeNet (Feature concatenation)	.922	.983	.965 - 1.000	.952	.885	.909	.939	.909	.952	.930	Train
GoogLeNet (Feature concatenation)	.781	.872	.803 - .942	.808	.732	.843	.682	.843	.808	.825	Test
GoogLeNet (Stacking fusion)	.878	.948	.911 - .985	.889	.865	.889	.865	.889	.889	.889	Train
GoogLeNet (Stacking fusion)	.833	.902	.843 - .961	.849	.805	.886	.750	.886	.849	.867	Test
DenseNet121 (Feature concatenation)	.913	.966	.933 - .999	.937	.885	.908	.920	.908	.937	.922	Train
DenseNet121 (Feature concatenation)	.798	.895	.839 - .951	.767	.854	.903	.673	.903	.767	.830	Test
DenseNet121 (Stacking fusion)	.878	.946	.903 - .989	.889	.865	.889	.865	.889	.889	.889	Train
DenseNet121 (Stacking fusion)	.789	.891	.832 - .950	.767	.829	.889	.667	.889	.767	.824	Test
VGG16 (Feature concatenation)	.939	.990	.975 - 1.000	.968	.904	.924	.959	.924	.968	.946	Train
VGG16 (Feature concatenation)	.754	.884	.825 - .942	.712	.829	.881	.618	.881	.712	.788	Test
VGG16 (Stacking fusion)	.878	.951	.915 - .987	.889	.865	.889	.865	.889	.889	.889	Train
VGG16 (Stacking fusion)	.781	.883	.825 - .942	.740	.854	.900	.648	.900	.740	.812	Test
MobileNetV3-Large (Feature concatenation)	.983	.991	.976 - 1.000	1.000	.962	.969	1.000	.969	1.000	.984	Train
MobileNetV3-Large (Feature concatenation)	.772	.887	.828 - .946	.767	.780	.862	.653	.862	.767	.812	Test
MobileNetV3-Large (Stacking fusion)	.913	.967	.930 - 1.000	.921	.904	.921	.904	.921	.921	.921	Train
MobileNetV3-Large (Stacking fusion)	.798	.880	.812 - .947	.795	.805	.879	.688	.879	.795	.835	Test

ACC, Accuracy; AUC, Area under receiver operating characteristic curve; CI, Confidence interval; F1, F1-score; NPV, Negative predictive value; PPV, Positive predictive value; PRE, Precision; SEN, Sensitivity; SPE, Specificity.

Figure 4.

ROC curves by fusion strategy. ROC curves comparing the performance of models with different fusion strategies in the training and test sets. (A-D) Feature concatenation fusion; (E-H) Stacking fusion. ROC, Receiver operating characteristic; AUC, Area under the receiver operating characteristic curve.

Evaluation of Multi-Parametric DL Features

The GoogLeNet FDLR signature obtained the best results by combining the signatures of the T1CE, T1WI, and T2WI (Table 4 and Figure 5). The results, including the AUC values for the combination, were significantly better than with the two-parameter (T1CE-T1WI) and single-parameter (T1CE, T1WI, and T2WI) models (P < .05) (Supplementary Table S5). Similarly, the AUC values of the two-parameter models were better than those of the single-parameter models. The confusion matrices for each model are shown in Supplementary Figure S6.

Table 4.

GoogLeNet FDLR Signatures Performance of Different Parameters in the Training and Test Sets.

Model name	ACC	AUC	95% CI	SEN	SPE	PPV	NPV	PRE	Recall	F1	SET
T1CE	.748	.835	.764 - .907	.810	.673	.750	.745	.750	.810	.779	Train
T1CE	.693	.753	.661 - .845	.726	.634	.779	.565	.779	.726	.752	Test
T1WI	.757	.796	.712 - .880	.794	.712	.769	.740	.769	.794	.781	Train
T1WI	.702	.768	.678 - .858	.712	.683	.800	.571	.800	.712	.754	Test
T2WI	.774	.864	.800 - .929	.825	.712	.776	.771	.776	.825	.800	Train
T2WI	.737	.823	.749 - .898	.740	.732	.831	.612	.831	.740	.783	Test
T1CE-T1WI	.809	.896	.840 - .952	.857	.750	.806	.813	.806	.857	.831	Train
T1CE-T1WI	.754	.836	.761 - .912	.753	.756	.846	.633	.846	.753	.797	Test
T1CE-T2WI	.809	.907	.854 - .960	.825	.788	.825	.788	.825	.825	.825	Train
T1CE-T2WI	.772	.868	.800 - .936	.781	.756	.851	.660	.851	.781	.814	Test
T1WI-T2WI	.843	.924	.879 - .969	.905	.769	.826	.870	.826	.905	.864	Train
T1WI-T2WI	.772	.871	.806 - .936	.795	.732	.841	.667	.841	.795	.817	Test
T1CE-T1WI-T2WI	.878	.948	.911 - .985	.889	.865	.889	.865	.889	.889	.889	Train
T1CE-T1WI-T2WI	.833	.902	.843 - .961	.849	.805	.886	.750	.886	.849	.867	Test

ACC, Accuracy; AUC, Area under receiver operating characteristic curve; CI, Confidence interval; F1, F1-score; FDLR, Fusion deep learning radiomics; NPV, Negative predictive value; PPV, Positive predictive value; PRE, Precision; SEN, Sensitivity; SPE, Specificity.

Figure 5.

ROC curves by variable parameters. ROC curves comparing the performance of models with different parameters in the training and test sets. (a) T1CE; (b) T1WI; (c) T2WI; (d) T1CE-T1WI; (e) T1CE-T2WI; (f) T1WI-T2WI; (g) T1CE-T1WI-T2WI. ROC, Receiver operating characteristic; AUC, Area under the receiver operating characteristic curve; T1CE, contrast-enhanced T1-weighted image; T1W1, T1-weighted image; T2W1, T2-weighted image.

Discussion

In present study, we evaluated FDLR signature based on multi-parametric MRI to predict TERT promoter mutation status in GBM patients. In the T2WI, T1CE, and T1WI, the corresponding GoogLeNet DLR signatures with 2 mm, 8 mm, and 10 mm expansion distances outside the tumour core through stacking fusion showed good discriminative power in both the training (AUC = .948), and external test (AUC = .902) sets. In addition, the high sensitivity and specificity of the model test set for identifying TERT promoter mutation might benefit GBM patients by allowing individualized surgical plans.

In comparing the single-parameter and single-region DLR models, this study found that DLR signatures based on the DL features of the peritumoural region were better than those of the intratumoural region (tumour core) in determining TERT promoter mutation status. Peritumoural features have also shown adequate performances in other medical image-based diagnostic tasks. For instance, one study used radiomics features of the peritumoural region to predict lymph node metastasis in early non-small cell lung cancer and achieved good classification (AUC = .825),²⁴ whereas another study used stratified peritumoural features to construct a nomogram and achieved good results (AUC = .804) in predicting lymph node metastasis in early cervical cancer.¹¹ It is well known that the depth of GBM invasion often extends beyond the tumour contrast enhancement area into the peritumoural region.¹⁸ Skogen et al.²⁵ used diffusion tensor imaging-based texture analysis to quantitatively analyze and evaluate the heterogeneity of tumour tissues and peritumoural edema in GBM and metastatic tumours. Whereas GBM and metastatic tumours showed no significant difference in solid tumour tissue heterogeneity, metastatic tumours had significantly greater peritumoural heterogeneity.²⁵ Therefore, the peritumoural area of GBM can provide valuable information about the tumour's biological characteristics.

Most models constructed by extracting DL features from four 2D CNNs models achieved valid results in predicting the presence of TERT promoter mutation. The discriminative power of predictive models using DL features was higher than those using radiomics and clinical features in most cohorts,^26,27 indicating that DL features incorporate additional information about important regions in the image, spatial tumour heterogeneity, and tumour microenvironment.²⁸ However, the first CNN training is prone to overfitting because the number of labelled medical images for a specific clinical problem is usually limited. Using pre-trained CNNs for feature extraction in the medical field has emerged as an effective way to address these difficulties.²⁹ Transfer learning can transfer prior knowledge of image features and apply it to novel medical images with better generalization, as verified in the present study.¹⁴

Furthermore, the multi-parameter model outperformed the single parameter models in identifying the TERT promoter mutation in GBM patients. The models' performance improved as the number of MRI sequences increased. A previous study used T1WI, T1WI-T1CE, and T1WI-T1CE-T2WI for radiomics analysis and achieved AUCs of .722, .795, and .843, respectively.³⁰ Consistent with our findings, they found that the multimodal MRI-based model outperformed the bimodal and unimodal models potentially since multimodal or multi-parametric learning can aggregate information from multiple sources. With sufficient training data, the richer the modalities or parameters input, the more accurate the estimation of the representation space that can be achieved.³¹

This study has some limitations. First, we selected the 2D image of the largest cross-section of the tumour instead of the entire tumour (3D) to extract DL features. The differences between 2D and 3D images may have affected some DL features. Therefore, DL feature extraction modeling based on images of the entire tumour is warranted in future studies. Second, we conducted a modeling analysis of only routinely used MRI sequences. Diffusion-weighted imaging (DWI), the apparent diffusion coefficient, and dynamic contrast-enhanced MRI (DCE-MRI) all have the potential to provide functional glioma information.³² Therefore, adding DWI and DCE-MRI may further improve the model's accuracy. Third, the FDLR signature for identifying TERT promoter mutation was used to evaluate patient prognosis. However, the features included in the FDLR signature cannot distinguish between the high- and low-risk GBM groups. Therefore, further radiomic research on intra- and peritumoural features to predict patient prognosis is needed.

Conclusion

Our results suggest that there may be a role for the intra- and peritumoural DL features in the identification of the TERT promoter mutation in patients with GBM. We propose an FDLR signature based on the multi-parametric MRI intra- and peritumoural fusion that may help in the preoperative identification of the TERT promoter mutation in GBM and inform individualized patient treatment.

Supplemental Material

Supplemental Material - Peritumoral Radiomics for Identification of Telomerase Reverse Transcriptase Promoter Mutation in Patients With Glioblastoma Based on Preoperative MRI

Supplemental Material for Peritumoral Radiomics for Identification of Telomerase Reverse Transcriptase Promoter Mutation in Patients With Glioblastoma Based on Preoperative MRI by Hongbo Zhang, Beibei Zhou, Hanwen Zhang, Yuze Zhang, Yi Lei, and Biao Huang in Canadian Association of Radiologists Journal.

Footnotes

Acknowledgments

We appreciate Python technology provided by the Onekey AI platform.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study is supported by a grant from National Natural Science Foundation of China (Grant Number: 82071871) and Youth Exploration Fund of Shenzhen Health Economics Society of China (Grant Number: 202211).

ORCID iD

Hanwen Zhang

Supplemental Material

Supplemental material for this article is available online.

References

Ostrom

Price

Neff

, et al. CBTRUS statistical report: primary brain and other central nervous system tumors diagnosed in the United States in 2015-2019. Neuro Oncol. 2022;24(Suppl 5). v1-v95. doi:10.1093/neuonc/noac202

Chinot

Wick

Mason

, et al. Bevacizumab plus radiotherapy-temozolomide for newly diagnosed glioblastoma. N Engl J Med. 2014;370(8):709-722. doi:10.1056/NEJMoa1308345

Osborn

Louis

Poussaint

, et al. The 2021 World Health Organization classification of tumors of the central nervous system: what neuroradiologists need to know. AJNR Am J Neuroradiol. 2022;43(7):928-937. doi:10.3174/ajnr.A7462

Louis

Perry

Wesseling

, et al. The 2021 WHO classification of tumors of the central nervous system: a summary. Neuro Oncol. 2021;23(8):1231-1251. doi:10.1093/neuonc/noab106

Fujimoto

Arita

Satomi

, et al. TERT promoter mutation status is necessary and sufficient to diagnose IDH-wildtype diffuse astrocytic glioma with molecular features of glioblastoma. Acta Neuropathol. 2021;142(2):323-338. doi:10.1007/s00401-021-02337-9

Silantyev

Falzone

Libra

, et al. Current and future trends on diagnosis and prognosis of glioblastoma: from molecular biology to proteomics. Cells. 2019;8(8):863. doi:10.3390/cells8080863

Weller

van den Bent

Tonn

, et al. European Association for Neuro-Oncology (EANO) guideline on the diagnosis and treatment of adult astrocytic and oligodendroglial gliomas. Lancet Oncol. 2017;18(6):e315-e329. doi:10.1016/S1470-2045(17)30194-8

Eckel-Passow

Lachance

Molinaro

, et al. Glioma groups based on 1p/19q, IDH, and TERT promoter mutations in tumors. N Engl J Med. 2015;372(26):2499-2508. doi:10.1056/NEJMoa1407279

Molinaro

Hervey-Jumper

Morshed

, et al. Association of maximal extent of resection of contrast-enhanced and non-contrast-enhanced tumour with survival within molecular subgroups of patients with newly diagnosed glioblastoma. JAMA Oncol. 2020;6(4):495-503. doi:10.1001/jamaoncol.2019.6143

10.

Perez-Morales

Tunali

Stringfield

, et al. Peritumoral and intratumoral radiomic features predict survival outcomes among patients diagnosed in lung cancer screening. Sci Rep. 2020;10(1):10528. doi:10.1038/s41598-020-67378-8

11.

Shi

Dong

Jiang

, et al. MRI-based peritumoral radiomics analysis for preoperative prediction of lymph node metastasis in early-stage cervical cancer: a multi-center study. Magn Reson Imaging. 2022;88:1-8. doi:10.1016/j.mri.2021.12.008

12.

Yan

Zhang

, et al. Quantitative MRI-based radiomics for noninvasively predicting molecular subtypes and survival in glioma patients. NPJ Precis Oncol. 2021;5(1):72. doi:10.1038/s41698-021-00205-z

13.

Zhuo

Feng

Yang

, et al. Radiomics nomograms of tumors and peritumoral regions for the preoperative prediction of spread through air spaces in lung adenocarcinoma. Transl Oncol. 2020;13(10):100820. doi:10.1016/j.tranon.2020.100820

14.

Dai

Liu

, et al. Combining multiparametric MRI features-based transfer learning and clinical parameters: application of machine learning for the differentiation of uterine sarcomas from atypical leiomyomas. Eur Radiol. 2022;32(11):7988-7997. doi:10.1007/s00330-022-08783-7

15.

LeCun

Bengio

Hinton

. Deep learning. Nature. 2015;521(7553):436-444. doi:10.1038/nature14539

16.

van Griethuysen JJM Fedorov

Parmar

, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104-e107. doi:10.1158/0008-5472.CAN-17-0339

17.

Liu

Chen

, et al. Radiogenomics to characterize the immune-related prognostic signature associated with biological functions in glioblastoma. Eur Radiol. 2023;33(1):209-220. doi:10.1007/s00330-022-09012-x

18.

Lemee

Clavreul

Menei

. Intratumoral heterogeneity in glioblastoma: don't forget the peritumoral brain zone. Neuro Oncol. 2015;17(10):1322-1332. doi:10.1093/neuonc/nov119

19.

Kolakshyapati

Adhikari

Karlowee

, et al. Nonenhancing peritumoral hyperintense lesion on diffusion-weighted imaging in glioblastoma: a novel diagnostic and specific prognostic indicator. J Neurosurg. 2018;128(3):667-678. doi:10.3171/2016.10.JNS161694

20.

Simonyan

Zisserman

. Very deep convolutional networks for large-scale image recognition. Computer Science 2014. doi:10.48550/arXiv.1409.1556.

21.

Szegedy

Liu

Jia

, et al. Going Deeper with Convolutions. IEEE Computer Society; 2014.

22.

Huang

Liu

Van Der Maaten

, et al. Densely connected convolutional networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017:2261–2269. doi: 10.1109/cvpr.2017.243.

23.

Howard

Sandler

Chen

, et al. Searching for MobileNetV3. Paper presented at: 2019 IEEE/CVF International Conference on Computer Vision (ICCV); 27 October 2019; Seoul, Korea. 2020.

24.

Wang

Zhao

, et al.

Can peritumoral radiomics increase the efficiency of the prediction for lymph node metastasis in clinical stage T1 lung adenocarcinoma on CT?

Eur Radiol. 2019;29(11):6049-6058. doi:10.1007/s00330-019-06084-0

25.

Skogen

Schulz

Helseth

, et al. Texture analysis on diffusion tensor imaging: discriminating glioblastoma from single brain metastasis. Acta Radiol. 2019;60(3):356-366. doi:10.1177/0284185118780889

26.

Cui

Zhang

, et al. A CT-based deep learning radiomics nomogram for predicting the response to neoadjuvant chemotherapy in patients with locally advanced gastric cancer: a multicenter cohort study. EClinicalMedicine. 2022;46:101348. doi:10.1016/j.eclinm.2022.101348

27.

Fukuma

Yanagisawa

Kinoshita

, et al. Prediction of IDH and TERT promoter mutations in low-grade glioma from magnetic resonance images using a convolutional neural network. Sci Rep. 2019;9(1):20311. doi:10.1038/s41598-019-56767-3

28.

Xie

Yang

, et al. Computed tomography-based deep-learning prediction of neoadjuvant chemoradiotherapy treatment response in esophageal squamous cell carcinoma. Radiother Oncol. 2021;154:6-13. doi:10.1016/j.radonc.2020.09.014

29.

Han

Qin

Bay

, et al. Deep transfer learning and radiomics feature prediction of survival of patients with high-grade gliomas. AJNR Am J Neuroradiol. 2020;41(1):40-48. doi:10.3174/ajnr.A6365

30.

Xie

Yang

Zhao

, et al. A deep learning model combining multimodal radiomics, clinical and imaging features for differentiating ocular adnexal lymphoma from idiopathic orbital inflammation. Eur Radiol. 2022;32(10):6922-6932. doi:10.1007/s00330-022-08857-6

31.

Huang

Xue

, et al. What Makes Multi-Modal Learning Better than Single (Provably). Computer Science. 2021.

32.

Zhang

Lyu

, et al. DSC and DCE histogram analyses of glioma biomarkers, including IDH, MGMT, and TERT, on differentiation and survival. Acad Radiol. 2020;27(12):e263-e271. doi:10.1016/j.acra.2019.12.010

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.26 MB

Peritumoural Radiomics for Identification of Telomerase Reverse Transcriptase Promoter Mutation in Patients With Glioblastoma Based on Preoperative MRI

Abstract

Keywords

Introduction

Methods

Patients

MRI protocol and preprocessing

Tumour Segmentation

Extraction and Selection of DL Features

Differentiation of TERT Promoter Status With Different Signatures

Statistical Analysis

Results

Patient Characteristics

Evaluation of Intra- and Peritumoural Regions

Evaluation of Different Multi-Parametric Fusion Strategies

Evaluation of Multi-Parametric DL Features

Discussion

Conclusion

Supplemental Material

Supplemental Material - Peritumoral Radiomics for Identification of Telomerase Reverse Transcriptase Promoter Mutation in Patients With Glioblastoma Based on Preoperative MRI

Footnotes

Acknowledgments

Declaration of Conflicting Interests

Funding

ORCID iD

Supplemental Material

References

Supplementary Material