Can Machine Learning of Magnetic Resonance Imaging Textural Features Differentiate Intra- and Extra-Axial Brain Tumours? A Feasibility Study

Hong Kong J Radiol 2025;28:Epub 12 September 2025

 

 

¹ Department of Radiology, University of Ottawa, Ottawa, Canada

² Systems Integration and Architecture, The Ottawa Hospital, Ottawa, Canada

 

  Full Article in PDF

Intracranial tumours can pose a diagnostic challenge in clinical practice. Recent advancements in artificial intelligence in radiology may provide additional diagnostic information. Radiomics techniques such as texture analysis can reveal grey-level patterns beyond what is possible through expert human visual perception alone.[1] Numerous quantitative textural features can be derived, including simple statistics based on grey-level histograms, and higher-order features based on spatial relationships among pixels.[2] These radiomic features, derived from conventional magnetic resonance imaging (MRI) sequences, can be leveraged to train various machine learning (ML) models to detect and classify brain tumours, as well as predict prognosis and treatment response.[3] [4] [5]

 

Previous applications of radiomics in brain tumour research have aimed to improve diagnosis and post-treatment imaging of gliomas, including grading, distinguishing tumour progression from pseudoprogression, patient survival, and genetic expression.[3] [4] Other neuro-oncologic ML advances include differentiating gliomas from mimics such as meningiomas, pituitary tumours, and solitary metastases,[5] [6] diagnosis of paediatric tumours,[7] [8] and detecting metastases.[9] Radiomic approaches can offer more consistent results with good external validity, compared to the interrater variability of human readers.[6] More recently, deep learning models trained to detect brain metastases have shown advantages over classical ML in terms of lower false-positive rates albeit with greater training data requirements.[9]

 

Accurately determining whether a lesion arises from within the brain parenchyma (intra-axial, IA) or from the surrounding structures (extra-axial, EA) is crucial for diagnosis and treatment planning. This study aimed to determine whether ML of MRI radiomic features can differentiate between IA and EA locations.

 

This retrospective study performed at a single tertiary-care academic centre. Medical record review was conducted per the guidelines of the Institutional Review Board. We identified 92 consecutive adult patients (age >18 years) who underwent brain MRI for a newly diagnosed solitary brain lesion from January 2017 to December 2018. Patients with multiple lesions or prior surgery or chemo/radiotherapy were excluded. Data on age, sex, and final histopathology were recorded.

 

MRI of the brain was performed using 1.5 T (25 IA and 9 EA tumours; Siemens Magnetom Symphony, Siemens Medical Systems, Erlangen, Germany) or 3 T (45 IA and 13 EA tumours; Siemens Tim Trio or GE DISCOVERY MR 750w, GE Medical Systems, Milwaukee [WI], US) systems. Using a dedicated head coil, three-dimensional (3D) axial stacks of T1-weighted (T1WI) post-contrast and T2-weighted images (T2WI) were acquired with the following parameters: 1.5 T: T2WI (fast spin echo with fat saturation): TR/TE 3510/97, echo train length 9, flip angle 180°, section thickness 5.5 mm; T1W post-contrast (fast spoiled gradient echo with fat saturation): TR/TE 6.73/2.71, flip angle 15°, section thickness 1 mm. 3 T: T2WI (fast spin echo with fat saturation): TR/TE 6700/97, echo train length 18, flip angle 120°, section thickness 3 mm; T1WI post-contrast (fast spoiled gradient echo): TR/TE 8.48/3.21, flip angle 12°, section thickness 1 mm. Post-contrast T1WI images were acquired after hand injection of 0.1 mmol/kg of gadobutrol (Gadovist; Bayer Healthcare, Hong Kong, China), followed by a 10-20 mL saline flush, using a 4–5-minute delay before acquisition at 3 T and a 6-to-7-minute delay at 1.5 T. All images were reviewed using a Picture Archiving Computed System (PACS; Horizon Medical Imaging, McKesson Corporation, San Francisco [CA], US).

 

The IA or EA designation was confirmed by final histopathology following biopsy, extracted from the electronic medical records. The specimens were obtained by the neurosurgeons and analysed by the neuropathologists at our centre.

 

Tumour volumes of interest (VOI) were manually segmented on both T2WI and post-contrast T1WI using ImageJ version 1.52r (National Institutes of Health, US, https://imagej.net/) by a neuroradiology fellow, under supervision of a staff neuroradiologist with over 30 years of experience. VOI contours were subsequently submitted to a blinded medical imaging scientist (redacted/blinded for review) for texture analysis. Examples are shown in Figure 1.

 

Figure 1. Manual segmentation in two sample patients. (a) T1-weighted post-contrast and (b) T2-weighted magnetic resonance imaging (MRI) sequences showing an extra-axial mass arising from the left frontal falx. (c) T1-weighted post-contrast and (d) T2-weighted MRI sequences from a different patient with a right anterior temporal mass.

 

First- and second-order statistical textural features were computed for each VOI and MRI sequence using MaZda software (version 4.6.0; Institute of Electronics, Technical University of Lodz, Lodz, Poland).[10] First-order features included grey-level histogram mean, variance, skewness, kurtosis, and percentile values (1st to 99th). Second-order features included grey-level co-occurrence matrix (GLCM[11]) and run-length matrix (RLM[12]) features (11 GLCM and 5 RLM features per sequence) Before computing GLCM and RLM features, signal intensities were normalised between μ ± 3σ (where μ was the mean value of grey levels inside the VOI [or VOI subzone] and σ was the standard deviation) and decimated to 32 grey levels to minimise inter-scanner variability.[13] [14]

 

We used XGBoost,[15] an open-source ML algorithm, to train models on the textural features extracted from T2WI, T1WI post-contrast, and their combination. Hyperparameters were tuned using an 100-trial Bayesian optimisation experiment via GPyOpt[16] (http://github.com/SheffieldML/GPyOpt), guided by Gaussian process modelling and an exploration-exploitation heuristic. Each model was evaluated using stratified ten-fold cross-validation, repeated 10 times.[17] The SHAP (Shapley Additive exPlanations) framework[18] was used to estimate each feature’s relative importance, normalised to sum to 1.0.

 

Statistical analysis was performed using RStudio, an open-source software (version 1.3.1093; PBC, Boston [MA], US). Mann-Whitney U tests were used to compare IA and EA groups for each feature. Stepwise Holm-Bonferroni correction was applied for multiple comparisons.[19] Model performance was assessed using accuracy, the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and F1 score. ROC confidence intervals (95% CI) were calculated using 5000 bootstrap iterations, and AUC differences were assessed using DeLong’s method.[20]

 

A total of 92 lesions were analysed (70 IA and 22 EA). Table 1 summarises demographics and histopathological data. Glioblastomas (44%) and metastases (27%) were the most common diagnoses in the IA group, while meningiomas (68%) were the most common in the EA group. The patients aged from 23 to 84 years, with a mean age of 58.8 years. No significant difference was noted between groups (IA median, 62 years [interquartile range, 18] vs. EA median, 55 years [interquartile range, 10]; p = 0.09). Males were more common in the IA group (66% vs. 23%; p < 0.01).

 

Table 1. Demographics and histopathologies (n = 92).

 

Table 2 provides median and interquartile range values for individual 3D textural features. EA tumours showed higher histogram mean, kurtosis, and 10th and 50th percentiles on T1WI post-contrast, and lower skewness (p < 0.001 for all). On T2WI, only kurtosis was significantly higher in the EA group (p < 0.001), while the 90th and 99th percentiles were significantly lower (p = 0.002 and 0.001, respectively).

 

Table 2. Median values of 25 individual textural features used in machine learning classification.

 

Among the second-order features computed from T1W post-contrast images, GLCM correlation, sum of squares, sum variance, and sum entropy associated with EA tumours were significantly lower than those computed for the IA group (p = 0.002 for correlation and p < 0.001 for the rest; Table 2). The T1 GLCM sum average and difference variance were both significantly greater in EA tumours compared to the IA group (p = 0.001 and p < 0.001, respectively). Similar to T1 post-contrast, the GLCM correlation, sum of squares, and sum variance features evaluated from T2 images were significantly lower than those in the IA group (p < 0.001 for each). The T2 GLCM difference variance was significantly greater in EA tumours compared to the IA group (p < 0.001). Among the group differences assessed for RLM features, none was found to be significant after Holm–Bonferroni correction for multiple comparisons (Table 2).

 

The classification performance metrics for each of the three ML models are summarised in Table 3. Receiver operating characteristic curves and feature attribution scores are depicted in Figures 2, 3, and 4. The AUC for the identification of IA tumours was 0.91 (95% CI = 0.89-0.93; Figure 2a) for the model based on T1-weighted post-contrast MRI features, 0.81 (95% CI = 0.78-0.84; Figure 3a) based on T2-weighted MRI features, and 0.92 (95% CI = 0.90-0.94; Figure 4a) based on all features.

 

Table 3. Classification performance metrics for each magnetic resonance imaging textural feature of the machine learning model.

 

Figure 2. (a) Receiver operating characteristic curve showing the area under the curve (AUC) for the XGBoost model, and (b) SHAP (Shapley Additive exPlanations) feature importance plot for the same model trained on T1-weighted post-contrast magnetic resonance imaging textural features to identify intra-axial tumours. SHAP values are aggregated across samples and normalised to sum to 1.0.

 

Figure 3. (a) Receiver operating characteristic curve showing the area under the curve (AUC) for the XGBoost model, and (b) SHAP (Shapley Additive exPlanations) feature importance plot for the same model trained on T2-weighted magnetic resonance imaging textural features to identify intra-axial tumours. SHAP values are aggregated across samples and normalised to sum to 1.0.

 

Figure 4. (a) Receiver operating characteristic curve showing the area under the curve (AUC) for the XGBoost model, and (b) SHAP (Shapley Additive exPlanations) feature importance plot for the same model trained on combined T1-weighted post-contrast and T2-weighted magnetic resonance imaging textural features to identify intra-axial tumours. SHAP values are aggregated across samples and normalised to sum to 1.0.

 

Table 3 shows that the three models (T1WI, T2WI, and combined) yielded AUCs that were significantly greater than 0.5 (p < 0.0001 for each comparison). The model produced by the T2WI features alone resulted in an AUC that was significantly lower than the model produced by either the T1W post-contrast or combined sequences (p < 0.0001 for each comparison). The combined model AUC was not significantly greater than the T1 model alone (p = 0.135). All models had high sensitivity in identifying intra-axial tumours (>90% for each), but none achieved high specificity (39%-64%). Nevertheless, the models attained acceptable levels of accuracy (>80%) and precision (>88%).

 

The feature importance attribution profile associated with the model trained using T1W post-contrast MRI features reveals that grey-level kurtosis, skewness, sum entropy, 10th (histogram) percentile, and sum variance contributed two thirds of the total proportional importance score (0.67/1.00) for this model (Figure 2b). Sum variance and kurtosis also contributed strongly towards the total proportional importance of the T2-based model (0.40/1.00), with sum of squares, difference variance, and 1st percentile contributing an additional 0.36/1.00 towards the total proportional importance for the T2-based model (Figure 3b). In the combined model, the T1WI skewness, kurtosis, sum entropy, and 10th percentile were found to represent four of the five ‘most important’ features, contributing 0.50/1.00 of the total proportional importance (Figure 4b).

 

Classically described features of EA lesions include broad-based dural attachment, adjacent bony changes, formation of a cerebrospinal fluid cleft, deviation of pial vessels, and buckling of the grey-white junction.[21] The most common primary EA lesions include meningiomas, schwannomas, pituitary adenomas, and Rathke’s cleft cysts, all of which have characteristic locations and signal properties that can aid the radiologist’s diagnosis.[21] For example, pituitary adenomas always arise in the sella or suprasellar location and may be completely T1 hypointense, or may also contain areas of cystic change and haemorrhage, and demonstrate delayed enhancement. Conversely, primary IA lesions are dominated by gliomas and lymphoma, with gliomas showing an infiltrative T2/fluid attenuated inversion recovery hyperintense pattern of spread along white matter tracts, and lymphoma characterised by greater diffusion restriction of its solid component.

 

Nevertheless, lesions can be challenging to localise visually, especially if large and associated with mass effect or oedema. These cases may require advanced imaging such as MRI perfusion and spectroscopy. In certain morphologically complex cases, the tumour origin may not be known until after biopsy. Utilising all available data, including quantitative radiomic features, could potentially improve treatment planning and avoid unnecessary procedures. Another practical benefit of radiomic models, should they achieve parity with human readers, would be the ability to reduce the number of imaging sequences required (e.g., axial T2WI and post-contrast T1WI, as used in our study).

 

We analysed 3D radiomic features of brain tumours using a ML framework (XGBoost), which determined IA versus EA location with high accuracy and AUC. The textural features used in the present study have been well described.[2] The T1W post-contrast model achieved a classification performance comparable to the combined T2WI and T1WI post-contrast model, and both performed significantly better than the model trained on T2WI-based features alone.

 

Various studies have utilised ML for similar tasks with excellent results. In a meta-analysis of 29 ML studies in neuro-oncology focused on patient outcomes, tumour characterisation, and gene expression, Sarkiss and Germano[4] reported a pooled sensitivity ranging from 78% to 98%, specificity from 76% to 95%, and greater accuracy compared to conventional imaging analysis in predicting clinical outcomes such as survival, high-versus low-grade tumours, and future progression. Tetik et al[5] developed an automated deep learning model to distinguish among gliomas, meningiomas, and pituitary tumours, achieving over 88% on all performance metrics including sensitivity, specificity, and Matthews correlation coefficient. A direct comparison of two human readers, a traditional ML model and a deep ML in differentiating glioblastoma from solitary metastases yielded similar performance, with AUCs of 0.77 and 0.90 in human readers and 0.89 and 0.96 for the traditional and deep ML models, respectively.[6] This study also highlighted the potential for robust generalisability with validation performed at a different institution, and greater inter-rater agreement between the ML models than the two human readers (albeit with differing levels of experience).[6]

 

Our study has several limitations. While all three models demonstrated high sensitivity for identifying IA tumours (>90% for each), none achieved high specificity (T1: 61%, T2: 39%, and combined: 64%). This is most likely due to the case imbalance between the IA (n = 70) and EA (n = 22) groups, which can artificially raise the sensitivity and accuracy for predicting IA tumours. The imbalance reflects the incidence of these tumours[21] and the nature of consecutive data acquisition. Second, the small sample size clearly limits the generalisability of our models. The current study was designed to assess whether MRI textural features contain sufficient predictive information for our XGBoost models to generate effective classifiers. Accordingly, the use of an established ten-fold cross validation method[17] was appropriate for estimating the generalisation errors of the models trained on T1WI post-contrast, T2WI, and combined MRI features. This preliminary stage is distinct from the development and validation of a single model intended for clinical deployment, which would require a much larger dataset and evaluation on external or ‘out-of-distribution’ data.[22] With improved optimisation, increased sample size and external validation, specificity could be enhanced and such a model could theoretically augment or complement the radiologist’s assessment in ambivalent cases. Finally, it should be noted that our dataset may be skewed towards aggressive or large lesions requiring resection, as we used a pathological reference standard. As a result, the extracted features could be skewed towards tumours that required resection, rather than asymptomatic lesions such as non-aggressive meningiomas.

 

A location-based ML classification model for differentiating IA from EA tumours is feasible based on this preliminary study, demonstrating good sensitivity. However, specificity was low to moderate, likely due to the imbalanced dataset. Further study with a more balanced cohort and external validation is required to optimise performance.