Magnetic resonance imaging for the differential diagnosis of primary extra-axial brain tumors: a review of radiomic studies

Full Text

BACKGROUND

Preoperative differential diagnosis of primary extra-axial brain tumors (PEABTs) is based on the analysis of magnetic resonance imaging (MRI) semiotics, which most commonly includes a standard set of weighted images (WI), such as T2-WI, T1-WI, FLAIR, diffusion-weighted imaging (DWI), and contrast-enhanced T1-WI (T1-CE) [1–3].

PEABTs include both benign and malignant neoplasms of the meninges (meningiomas and mesenchymal tumors) and cranial nerves (neurinomas) [4].

The MRI semiotics of PEABTs have been studied in detail and described in established guidelines; however, atypical MRI patterns can complicate the differential diagnosis of tumors based on visual assessment alone [5, 6]. Incorrect tumor type determination can result in incorrect treatment [1, 2, 7, 8]. The most common difficulties are differentiating meningiomas of various grades, distinguishing solitary fibrous tumors from meningiomas, and localizing PEABTs in cerebellopontine angles [9–12].

Radiomics is a quantitative approach to medical image analysis and aims to identify the relationship between the digital characteristics of a diagnostic image and phenotypic and genotypic characteristics of a tumor [13].

Radiomics involves extracting quantitative features from images to provide an objective description of an imaging phenotype and determine the relationship between radiomic and genetic, molecular, and clinical features of tumors [14]. To extract quantitative parameters from images, morphometric, histogram, and texture analysis of segmented areas of interest is performed. Histogram and texture features reflect structural features not detectable visually [15]. In radiomics studies, various mathematical modeling and deep learning methods are used. The resulting differential diagnostic and prognostic models should be validated using an independent sample. Radiomics may be a powerful tool in clinical decision-making [16]. Figure 1 shows the steps of radiomic analysis.

Fig. 1. Radiomic analysis stages.

Some analytical publications have summarized previous studies on radiomics-based differential diagnosis of PEABTs [13, 17]. The rapid accumulation of new clinical cases and increase in the number of studies related to this problem require further analysis and systematization, and thus, is the basis of the present study.

MATERIALS AND METHODS

A systematic review for the last 5 years was conducted according to PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Protocols, 2009).

PubMed/MEDLINE, Google Scholar, and eLibrary databases were searched for scientific articles from the last 5 years using the following keywords: MRI, meningioma, neuroma, schwannoma, solitary fibrous tumor, radiomics, texture, МРТ, менингиома, невринома, шваннома, солитарная фиброзная опухоль, радиомика, текстура.

Articles with abstracts unrelated to the differential diagnosis of PEABTs were excluded, as well as those without a description in text of radiomic features in the differential diagnosis of tumors.

Finally, 19 publications were included in the review. Figure 2 shows the design of the current study.

Fig. 2. Research design.

Estimated parameters

As part of the systematic review, the following parameters were assessed in the selection of publications:

• Diagnostic task
• Number of patients
• Method of tumor segmentation
• WI types
• Significant differential diagnostic features
• Mathematical models used
• Validating the mathematical models used to classify tumors

This review included data from original clinical trials.

RESULTS

Several studies have investigated the effectiveness of radiomic features of MRI images in the differential diagnosis of PEABTs. In total, 121 studies published in the last 5 years were found in PubMed/MEDLINE and Google Scholar for the search query “meningioma or neuroma or schwannoma or solitary fibrous tumor) + (texture or radiomic) + MRI.” For queries with different combinations of the words “MRI, meningioma, neuroma, schwannoma, solitary fibrous tumor, radiomics, and texture,” only one publication was found in eLibrary. After analyzing the publications, 19 articles in English and Russian were included in the review.

Table 1 shows the characteristics of the selected articles according to research design.

Table 1. Diagnostic tasks of radiomic analysis for the differential diagnosis of primary extra-axial brain tumors

Note. Abbreviations: Tumors: Mb, benign meningiomas (grade 1); Mm, malignant meningiomas (grades 2 and 3); М, meningiomas without grade; N, neurinomas; C, craniopharyngiomas; H, hemangiomas; SFT, solitary fibrous tumors; Gr, grade Segmentation: Man, manual; SA, semiautomatic; A, automatic Features: ADC, apparent diffusion coefficient; SWI, susceptibility weighted imaging; Histo, histogram; GLCM, gray level co-occurrence matrix; GLRLM, gray level run length matrix; GLSZM, gray level size zone matrix; NGTDM, gray-tone difference matrix; AU, autoregressive model; Wav, wavelet Mathematical modeling methods: LR, logistic regression; NB, naive Bayes classifier; SVM, support vector machine; TC, text categorization; KNN, K-nearest neighbors method; DT, decision tree; RF, random forest; LDA, linear discriminant analysis; XGB, extreme gradient boost; MLP, multilayer perceptron Test information parameters: Acc, accuracy; Sn, sensitivity; Sp, specificity; AUC, area under the curve Other: NP, not performed; NA, not available

DISCUSSION

Differential diagnosis

Twelve studies were found to have investigated the differential diagnosis of benign and malignant meningiomas. The differential diagnosis of meningiomas and solitary fibrous tumors was evaluated in four studies, and the differentiation between meningioma and hemangioma/craniopharyngioma/neurinoma was examined in one study.

Most studies have discussed a “binary” classification between two types of PEABTs [18, 19, 22-32, 34, 35]. Given the similar semiotics of all PEABTs, models capable of performing multiclass rather than binary classification between two prespecified tumor types have an advantage for clinical use. However, only four studies have distinguished between three or more types of PEABTs [20, 21, 33, 36].

Comparison of the effectiveness of tumor classification by visual assessment and models based on mri semiotics or radiomic parameters

Two studies have compared the effectiveness of tumor classification between radiologists and radiomics models [20, 30]. In these studies, the tumor type was determined by a radiologist based on the MRI image, without mathematical modeling based on visual features. The use of radiomics models was advantageous, with an accuracy of 10%–17%.

Five studies have compared the accuracy of tumor classification between models based on MRI semiotics and radiomics features [20, 23, 32, 33, 36]. In these studies, the visual semiotic features were systematized and stratified. Based on these features, mathematical models may be developed.

The use of mathematical semiotic models for tumor classification may be more advantageous over the radiologist’s opinion because a radiologist’s differential diagnosis is largely based on experience and subjective. Moreover, systematizing and integrating the evaluation of MRI semiotic features increases their information value.

Differentiating tumors using models based on radiomic parameters was significantly superior to classifying tumors by radiologists, and their information value was higher than that of semiotic models.

The most valuable studies are those that compare the information value of radiomics models with the results of visual assessment of MRI semiotics. Furthermore, the ability to automate image analysis for computer decision support systems remains an advantage of the radiomics approach.

Patient sample size and model validation

Most studies have included relatively small numbers of patients:

• <100 patients: 6 publications [19, 24, 27, 30, 35, 36],
• 100–200 patients: 7 publications [18, 20-22, 28, 31, 34],
• 200–300 patients: 4 publications [25, 29, 32, 33],
• >300 patients: 2 publications [23, 26].

Larger samples are typical for differential diagnosis studies of meningiomas. The small sample size may be because of the unequal prevalence of the different types of PEABTs. Most PEABTs (> 80%) are benign meningiomas, and other tumors are rare [4], making it challenging to select a large patient population.

Validation was completed in 84.2% of the trials. In most studies, the validated group comprised approximately one-third of the total enrolled patients. Note that the clinical significance of differential diagnostic models is reduced by the lack of testing of model performance on the validation set.

Tumor segmentation

Segmenting the tumor is the first and fundamental step in radiomics analysis [14]. To avoid distortion of radiomic features and ensure reproducibility of results, the segmentation technique should accurately distinguish neoplastic tissue from brain matter and peripheral edema.

Most PEABTs are characterized by a marked increase in MRI signal intensity on T1-WI after contrast administration, whereas the isointense and hypointense MRI signals of adjacent structures are preserved [9]. This feature is the basis for the sharp difference in brightness between the contrasted tumor and adjacent structures and accuracy of tumor margin visualization. Most studies included in the review (63.2%) have performed segmentation specifically on contrast-enhanced T1-WI [18, 21-28, 30, 33, 36].

The segmentation method affects the final simulation result. The automatic and semiautomatic methods have a higher reproducibility than the manual methods in the determination of tumor boundaries [37]. In the presented studies, less preferred manual segmentation was most common [18-20, 22, 23, 27-29, 31-33]. Only five studies have used automated or semiautomated methods [18, 19, 26, 27, 36].

Significant radiomic features

A feature of radiomics studies is the presence of a sufficiently large initial set of parameters, and the most informative parameters are selected to solve the problem. Histogram and texture parameters of tumors are the most informative radiomic features for the differential diagnosis of PEABTs.

The power of radiomics models for the differential diagnosis of PEABTs is increased by expanding the set of radiomic parameters to include different WI types. Seven studies have compared models based on the radiomic parameters of one WI type with models that included the features of different WI types [19, 25, 28, 29, 31-33]. In six of these studies, the advantage of the latter was demonstrated based on a comparison of the information values of the tests [19, 25, 29, 31-33]. One study has revealed the poor results of weight combination models [28].

In comparing the two combined models, Hu et al. [26] have shown that a model containing an extended spectrum of weights (T2-WI, T1-WI, T1-CE, apparent diffusion coefficient [ADC] map, susceptibility weighted imaging [SWI]) was slightly superior to a model based on T2-WI, T1-WI, and T1-CE.

The advantage of models using several types of weights is their ability to reflect different aspects of the tumor. For example, T2-WI and T1-WI reflect the degree of hydration (amount of fluid) in the tumor, T1-CE reflects the permeability of the blood–brain barrier, DWI and ADC reflect the cellularity of the tumor, and SWI is sensitive to hemorrhage and fossilization. Therefore, integrating the parameters within the model allows a more complete representation of the morphological characteristics and better results.

The shape parameter values were limited. The information value of these parameters was evaluated in ten studies [21-23, 26-28, 31-33, 35]. Shape parameters were informative in studies that have performed modeling based on one WI type [21-23, 27, 28, 35]. Three studies [31-33] have shown that shape parameters are uninformative when constructing models that include multiple WI types. In a study by Hu et al. [26], shape parameters were informative and were included in the modeling; however, their proportion was not large compared with that of histogram and texture parameters (the model included 17 histogram and texture parameters and 3 morphometric parameters).

Mathematical modeling methods

In the presented studies, various mathematical modeling methods were used to create models. The most common algorithms were as follows:

• Support vector machine (SVM): 9 studies [18, 21, 22, 25, 28, 30, 32, 33, 35],
• Logistic regression (LR): 10 studies [19, 24, 27-29, 31-35],
• Random forest (RF): 6 studies [18, 23, 26, 28, 33, 35].

Three studies [18, 28, 33] have analyzed the results of tumor classification using models based on these methods and showed conflicting results. In a study by Park et al. [18] (RF and SVM) and in another by Wei et al. [33] (LR, RF, SVM), the methods showed a comparable level of information value. However, in a study by Han et al. [28], the results varied significantly according to the modeling technique (one of the information value parameters of the test, area under the curve (AUC), varied from 0.628 to 0.922), whereas the SVM showed more stable results.

Among all modeling methods, the best information value parameters were demonstrated by LR [27] and SVM [35], wherein the AUC was 0.95.

CONCLUSION

The use of radiomics approach shows promising results in the differential diagnosis of PEABTs. Additionally, clinical practice implementation requires greater methodological rigor in the conduct of radiomics studies, including mandatory validation, standardization of segmentation methods, determination of the required feature set, and more informed choice of mathematical modeling methods. The use of histograms and texture parameters of different WI types for further revealing the potential of radiomics in the differential detection of PEABTs appears favorable.

Prospective studies using automated segmentation methods and an expanded set of WI types and the development of radiomics models that allow multiclass differential diagnosis of PEABTs may lay the foundation for creating powerful tools for digital clinical decision support systems and can ensure optimal patient treatment selection.

ADDITIONAL INFORMATION

Funding source. This study was not supported by any external sources of funding.

Competing interests. The authors declare that they have no competing interests.

Authors’ contribution. All authors made a substantial contribution to the conception of the work, acquisition, analysis, interpretation of data for the work, drafting and revising the work, final approval of the version to be published and agree to be accountable for all aspects of the work. The major contributions were distributed as follows: А.V. Kapishnikov — the concept of the study, approval of the final version; E.N. Surovcev — concept and design of the work, manuscript text writing and editing, collection and processing of materials, data analysis.

About the authors

Aleksandr V. Kapishnikov

Email: a.v.kapishnikov@samsmu.ru
ORCID iD: 0000-0002-6858-372X
SPIN-code: 6213-7455
Scopus Author ID: 6507900025

MD, Dr. Sci. (Med.), Professor

Evgeniy N. Surovcev

Author for correspondence.
Email: evgeniisurovcev@mail.ru
ORCID iD: 0000-0002-8236-833X
SPIN-code: 5252-5661
Scopus Author ID: 57224906215
Russian Federation, Samara; Togliatti

References

Supplementary files

Supplementary Files

Action

1. JATS XML

Download

2. Fig. 1. Radiomic analysis stages.

Download (265KB)

Indexing metadata

3. Fig. 2. Research design.

Download (230KB)

Indexing metadata