Utilizing the amide proton transfer technique to characterize diffuse gliomas based on the WHO 2021 classification of CNS tumors

Abstract

Purpose

Diffuse gliomas present a significant challenge for healthcare systems globally. While brain MRI plays a vital role in diagnosis, prognosis, and treatment monitoring, accurately characterizing gliomas using conventional MRI techniques alone is challenging. In this study, we explored the potential of utilizing the amide proton transfer (APT) technique to predict tumor grade and type based on the WHO 2021 Classification of CNS Tumors.

Methods

Forty-two adult patients with histopathologically confirmed brain gliomas were included in the study. They underwent 3T MRI imaging, which involved APT sequence. Multinomial and binary logistic regression models were employed to classify patients into clinically relevant groups based on MRI findings and demographic variables.

Results

We found that the best model for tumor grade classification included patient age along with APT values. The highest sensitivity (88%) was observed for Grade 4 tumors, while Grade 3 tumors showed the highest specificity (79%). For tumor type classification, our model incorporated four predictors: APT values, patient’s age, necrosis, and the presence of hemorrhage. The glioblastoma group had the highest sensitivity and specificity (87%), whereas balanced accuracy was the lowest for astrocytomas (0.73).

Conclusion

The APT technique shows great potential for noninvasive evaluation of diffuse gliomas. The changes in the classification of gliomas as per the WHO 2021 version of the CNS Tumor Classification did not affect its usefulness in predicting tumor grade or type.

Introduction

Diffuse glioma is the most common primary brain tumor, accounting for approximately 80% of all primary malignant brain tumors in adults. ¹ Overall survival and prognosis for relapse, as well as the provision of adjuvant therapeutic regimens, depend dramatically on both tumor grade and morphological subtype. ² In addition, it is believed that the molecular profile of the tumor can predict the response to treatment.^3,4 Thus, an accurate evaluation of the malignant potential of diffuse gliomas is necessary.

The last release of the WHO Classification of CNS Tumors (2021) is focused primarily on molecular diagnostics, where the main factors for grading and tumor type are the combination of IDH mutation and 1p19q codeletion statuses. ⁵ However, tumor grading is still carried out through visual analysis, ⁶ which is subjective and not always accurate due to tumor heterogeneity. ⁷ The evaluation of the Ki-67 proliferation index has been recognized as highly beneficial in achieving this objective. ⁸ However, it necessitates additional immunohistochemical staining. Consequently, despite morphological analysis being regarded as the definitive method for diagnosing diffuse gliomas, there is variability among observers, including experts. ⁹ Furthermore, detailed histological analysis with tumor molecular profile evaluation is expensive and requires advanced laboratory techniques, which could be a significant problem in middle- and low-income countries. ¹⁰

Brain magnetic resonance imaging (MRI) plays a key role in the diagnosis, presurgical planning, surveillance, and treatment monitoring of gliomas. ¹¹ To unify radiological reports and create a common vocabulary, the VASARI scoring system has been developed for a detailed description of brain gliomas (https://wiki.cancerimagingarchive.net/display/Public/VASARI+Research+Project), and its use has already shown promising results.^12,13 However, the signal changes on most conventional MRI sequences lack biological specificity, which limits the accuracy of noninvasive glioma characterization. Therefore, although conventional MRI is readily accessible and provides crucial anatomical details, accurately distinguishing the type and grade of a tumor solely based on conventional techniques appears to be challenging.^11,14

The advancements and widespread adoption of advanced MRI sequences in recent decades have enabled clinicians and researchers to gather extensive information about tumor structure and physiology, facilitating noninvasive glioma diagnosis and evaluation of treatment effectiveness. ¹¹ One of the promising MRI techniques used for imaging diffuse brain gliomas is amide proton transfer (APT), known for its remarkable biological specificity. It is well known that gliomas have higher protein/peptide contents than normal brain tissue. ¹⁵ Therefore, information at the protein level could potentially contribute to earlier diagnosis, more accurate delineation of boundaries, and enhanced tumor characterization. ¹⁶ The biophysical basis of APT imaging is the ability to detect mobile proteins. ¹⁷ The clinical utility of the technique seems promising, for example, it can be used in planning stereotactic biopsy. ¹⁸ In addition, it is a noninvasive procedure that does not require the administration of any contrast agents.

APT showed better diagnostic performance than conventional MRI. ¹⁹ Moreover, it has been discovered that it is just as effective as DSC perfusion. ²⁰ A recently published meta-analysis confirmed the effectiveness of the APT technique in distinguishing between low- and high-grade gliomas, as well as its potential for predicting histopathology noninvasively. ¹⁷ However, some authors have reported improved accuracy in grade prediction when combining APT with ADC and ASL values. ²¹ In addition, several studies^22,23 have found a correlation between APT values within the glioma and tumor cellularity. Furthermore, these studies have also shown a relationship between APT values and the extent of diffusion restriction as measured by both ADC and DKI techniques.^22,24

Beyond that, APT seems to be a predictive factor for IDH mutation status as well.^18,25,26 It is important to note that the application of APT is not limited to presurgical glioma diagnosis and can also have implications for predicting overall survival, prognosis of recurrence, and assessing treatment outcomes. This has been demonstrated in multiple recently published articles.^26–28

As a reflection of its advantages and benefits for clinical applications, numerous articles focusing on the radiological characteristics of diffuse gliomas have been published over the past years. Interestingly, only a few authors have tried to determine the subtype of glioma, despite its clear correlation with patient prognosis. However, it is worth noting that most of these studies were conducted using the WHO 2016 Classification of CNS Tumors, whereas the recently released WHO 2021 update introduces significant revisions to glioma grading. Several published articles on neuroimaging of glioma used the WHO 2021 classification.^29,30 However, we were unable to find published data dedicated to the application of APT technique based on new classification.

Thus, the main objective of this study was to investigate the usefulness of the APT technique in predicting the grade and tumor type according to the WHO 2021 Classification of CNS Tumors for diffuse gliomas. Furthermore, we sought to assess the ability of APT (alone or combined with other variables) to determine the status of IDH mutations, which greatly influences prognosis. Finally, we were interested in examining the relationship between quantitative MRI data and tumor cellularity measured by Ki-67 levels.

Methods

Patients

The subjects in this prospective study were patients with first identified brain gliomas who were surgically treated at our hospital from January 2023 to August 2023. Forty-two patients (20 males and 22 females, 22–76 years of age) with morphologically proven gliomas (according to WHO 2021 criteria) participated in the study. None of the patients had previously undergone neurosurgical treatment. All patients underwent high-resolution brain MRI before surgery. Each patient signed a written informed consent form to participate in the study. The study was carried out according to the Declaration of Helsinki and was approved by the local Ethics Committee of the Federal Center for Neurosurgery, Novosibirsk, Russia (protocol No. 4 dated 02-08-2022). Detailed information on the patients is provided in Supplementary Table 1.

Morphology

The surgical samples were assessed using the 2021 WHO classification of CNS tumors. ⁵ Immunohistochemical staining was conducted to evaluate the IDH1 status in each case. Additionally, FISH analysis was performed to detect 1p19q codeletion. Ki-67 levels were assessed using immunohistochemistry.

Magnetic resonance imaging acquisition

MR imaging data were acquired using a 3T system (Ingenia, Philips Healthcare, The Netherlands) equipped with a 16-channel receiver head coil. The MRI protocol included high-resolution T1-WI (before and after contrast injection), T2-WI, FLAIR, SWI, DWI, ASL, and APT sequences. Details regarding the acquisition parameters used are shown in Table 1.

Table 1.

Parameters of MRI sequences.

	T1WI/CE-T1WI	T2WI	FLAIR	APT	ASL	ADC	SWI
Sequence type	3D TFE	3D TSE	3D TSE	3D SE	3D pCASL	2D EPI	3D GRE
TR (msec)	6.7	3500	4800	5925	4300	3800	31
TE (msec)	2.98	300	340	8.3	11.6	90	0
Flip angle	8	90	90	90	90	90	17
Matrix	256 × 256	256 × 256	228 × 228	128 × 128	64 × 64	128 ×128	384 × 255
FOV (mm)	256	256	256	256	240	256	256
Slice thickness (mm)	1	1	1.2	6	6	4	2.4
N of slices	192	360	140	10	14	26	50
Inversion time (msec)	/	/	1650	/	/	/	/
b values	/	/	/	/	/	0, 500, 1000	/
PLD (msec)	/	/	/	/	1800	/	/
Acquisition time (min:sec)	6:11	5:19	4:24	4:30	5:01	2:05	1:20

Magnetic resonance imaging data processing

The quantitative APT maps were calculated automatically by the MRI system. In each case, the entire tumor was segmented with ITK-Snap software (version 4.0.0, https://www.itksnap.org) using a semiautomatic classification algorithm. For patients with contrast-enhancing gliomas, segmentation was performed based on postcontrast T1-WI sequence (referring to FLAIR, T2-WI). For patients with non-contrast-enhancing gliomas, segmentation was performed based on FLAIR and T2-WI sequences (because the contrast between tumor and brain tissue is highest in these modalities). The segmentation results were saved as binary masks in NifTI format. The APT maps were registered and resampled referring to T1-WI using SPM12 with a normalized mutual information cost function and 4th Degree B-Spline interpolation (https://www.fil.ion.ucl.ac.uk/spm/). Subsequently, the tumor binary mask was moved to the APT maps, and quantitative analysis was performed using Pyradiomics tool (https://aim.hms.harvard.edu/pyradiomics). The following signal intensity characteristics were extracted from the tumor on the APT maps: (1) mean, (2) median, (3) 90th percentile. Absolute values of all quantitative MR parameters were used. An illustration of data processing is shown in Figure 1.

Figure 1.

An example of post-processing of MRI data in a patient with glioblastoma (WHO grade 4, IDH1 wild type). A—Axial CE-T1WI shows vivid inhomogeneous enhancement; B—same slice after semiautomatic tumor segmentation; C—axial T2WI; D—APT map demonstrates marked elevation of metabolites in the tumor’s center; E—CBF map (ASL) with slight hyperperfusion from the tumor; F—ADC map shows moderate diffusion restriction.

Two neuroradiologists qualitatively evaluated the MRI data independently (with 5 and 2 years of neuroradiology experience, respectively). Magnetic resonance features were defined according to Visually Accessible Rembrandt Images (VASARI) imaging criteria (https://wiki.cancerimagingarchive.net/display/Public/VASARI+Research+Project), and full results are provided in Supplementary Table 2. Subsequently, qualitative features such as enhancement quality (f4), necrosis proportion (f7), and hemorrhage presence (f16) were evaluated between groups.

Statistical analysis

Descriptive statistics are presented as the median (interquartile range, IQR). The relationship between measured variables was assessed with the Spearman correlation coefficient and FDR correction for multiple comparisons. Ninety-five percent confidence intervals are provided in square brackets following the correlation coefficient magnitude. The chi-squared test was used for categorical data analysis. The Mann‒Whitney test was utilized for comparison of metric variables grouped by one categorical variable, such as sex, IDH1 mutation and 1p19q codeletion (two levels). Alternatively, in the case of categorical variables such as grade or tumor type, both having three levels, the Kruskal‒Wallis test was used instead. A post hoc Dunn test was implemented to estimate the statistical significance between the studied groups in a pairwise manner. To assess the ability of the measured variables to classify patients into the abovementioned clinically meaningful groups (grade, tumor type, IDH1 mutation and 1p19q codeletion), we ran a number of binomial and multinomial logistic regression models. For each model reported in the paper, we provided cross-validated (training/test sample—70/30) values of accuracy, sensitivity, specificity and area under the curve (AUC). In the case of tumor grade prediction, second grade was used as a reference category. For the tumor type prediction, we chose astrocytoma as a reference. In addition, we used a likelihood ratio (LR) test to assess the difference between distinct versions of nested models, each containing a different number of predictors. Odds ratios and associated 95% confidence intervals were additionally calculated for all predictors included in the final versions of the models. The values for factors f4, f7, and f16 were obtained through their independent assessment by two neuroradiologists and subsequent calculation of the intraclass correlation coefficient based on a mean-rating (k = 2), absolute-agreement, 2-way mixed-effects model. A p value of 0.05 was considered a threshold for evaluating statistically significant associations. All statistical analyses were run in R (v. 4.3.1, 2023).

Results

Patient characteristics

A total of 42 patients (20 males, 22 females) with brain gliomas participated in the study. There were 15 patients with glioblastomas, 18 patients with astrocytomas (3 with grade 2, 7 with grade 3, and 8 with grade 4), and 9 patients with oligodendrogliomas (2 with grade 2 and 7 with grade 3). The median age of the recruited patients was 54.5 years (IQR = 22.75). Male and female patients did not differ in age (U = 267.5, p = .24), grade (χ² = 3.29, df = 2, p = .19) or 1p19q codeletion presence (χ² = 0.84, df = 1, p = .36). Detailed information on demographic, morphology, and molecular data for each patient is provided in Supplementary Table 1. Neurosurgical treatment was performed in each case. There were 26 cases with gross total tumor resection, 10 cases with subtotal resection, 4 cases with partial resection, and 2 cases with open biopsy.

Qualitative MRI features

The full results of the qualitative MRI analysis according to VASARI are demonstrated in Supplementary Table 2. Three characteristics, quality of enhancement (f4), necrosis (f7), and presence of hemorrhage (f16), were chosen for the following analysis. When evaluating the agreement of results obtained from comparing the levels of factors f4, f7 and f16 between two neuroradiologists, we observed intraclass correlation coefficient values of 1 for factors f4 and f7, while factor f16 yielded a value of 0.7, indicating a high degree of consistency in estimates between the raters. As a next step, a chi-square goodness of fit test was performed to determine whether the proportions of these variables were equal between groups of patients with different tumor types and grades. There were significant relations between tumor types and the presence of hemorrhage (χ² = 7.1, df = 2, p = .029), necrosis (χ² = 21.4, df = 6, p = .0016) and quality of enhancement (χ² = 23.28, df = 4, p = .0001). Along the same lines, these variables also differed between patients with different tumor grades (χ² = 8.18, df = 2, p = .017 for hemorrhage; χ² = 19.38, df = 6, p = .0036 for necrosis; χ² = 21.9, df = 4, p = .0002 for enhancement).

Quantitative MRI features

A Kruskal‒Wallis test was performed on the median APT values of the three groups (grades 2, 3, and 4) and revealed that there was a statistically significant difference in these values among the tested groups (χ² = 11.25, df = 2, p = .0036). A post hoc Dunn test was further conducted to determine the differences in measured values between each pair of groups. We found the abovementioned differences in grade 2-grade 4 and grade 3-grade 4 pairs (p = .03 and p = .01, respectively; Figure 2(A)).

Figure 2.

The results of nonparametric statistical comparison of APT median values between groups with Kruskal‒Wallis (panels A and B) and Mann‒Whitney tests (panels C and D). APT levels are compared among groups using boxplots, where individual data values are represented by black dots and the medians are shown by the thick red horizontal lines. The boxes represent the middle 50% of the data, ranging from the 25th to the 75th percentile. The whiskers extend to 1.5 times the interquartile range. In this figure, NS indicates p > .05; * represents p < .05; **p < .01; ***p < .001.

In a similar vein, median APT values were compared between patients belonging to groups with different tumor types, resulting in a rejection of the null hypothesis (χ² = 17.33, df = 2, p = .0002). Patients with glioblastoma showed increased median APT values in comparison to those with astrocytoma and oligodendroglioma (p = .002 and p = .0004, respectively; Figure 2(B)). Additionally, we found that APT values were higher in patients with glioblastoma than in those with astrocytoma grade 4 (U = 110, p = .01).

When comparing APT median levels between patients with IDH1 mutant and wild types, the latter group demonstrated increased median values, which were statistically significant (U = 51, p = 7.37*10⁻⁵; Figure 2(C)). Similarly, when applied to groups with and without 1p19q codeletion, the Mann‒Whitney test indicated the presence of noticeable differences in median APT levels (U = 236, p = .0077; Figure 2(D)). Differences in the mean and 90th percentile APT values were also statistically significant between groups, but these differences were less pronounced than those of median APT. We report these results in detail in Supplementary Table 3.

Tumor grade and type prediction

To predict a patient’s tumor grade based on a linear combination of MRI parameters as well as clinical and demographic variables, we ran several multinomial logistic regression models, characterized by different numbers of predictors or independent variables. The LR test showed that comparison of the null model and the updated version with the addition of APT values returned statistically significant differences between models (χ² = 13.17, df = 2, p = .0014). Including the age variable in the model resulted in further improvement in its performance and led to differences from the model that used APT values as a single predictor (χ² = 8.7, df = 2, p = .012). Entering additional variables, such as necrosis, hemorrhage, or quality of enhancement, into the model did not result in any significant model improvement, so we did not report them in the article. The best model containing APT and age as the main predictors had a multiclass AUC of 0.82 and an accuracy of 0.71 [0.35, 0.94]. The balanced accuracy values were 0.5 for grade 2, 0.74 for grade 3, and 0.77 for grade 4. Notably, grade 2 showed a sensitivity of 0% but a specificity of 100%. Grade 4 demonstrated the highest sensitivity (88%) among all groups tested while having a moderate magnitude of specificity (67%). Finally, the model was able to correctly classify patients as belonging to grade 3 in 69% of cases, with a specificity value reaching 79% (Figure 3(A)). Table 2 shows the logs of the odds ratio for each predictor included in the model, 95% confidence interval and corresponding p value.

Figure 3.

The plot represents areas under the curves (ROC curves) for two multinomial logistic regression models (one-vs-the-rest multiclass strategy) used to classify patients according to their tumor grade (panel A) or tumor type (panel B). Different grades or types are depicted in different colors. The dotted line corresponds to AUC = 0.5 or the random classifier’s performance.

Table 2.

Parameters of the best (two-predictors) model, classifying patients into three tumor grades. Reference category was chosen to be the second grade. Table provides the odds ratio, 95% confidence interval and corresponding p-values for each predictor in the model. OR–odds ratio, CI—confidence interval.

Characteristic	Third grade	Fourth grade	p-value	OR	95% CI	p-value
	OR	95% CI
Median APT	1.33	0.24, 7.32	0.7	8.52	1.28, 56.5	0.026*
Age	0.96	0.89, 1.04	0.3	1.05	0.97, 1.15	0.2

*statistical significance at the level of p < .05

Similarly, with the objective of classifying patients into distinct tumor types, we constructed multiple multinomial logistic regression models with varying numbers of predictors. Initially, we included APT median values as a single predictor, following prior grade prediction models, and found that this model exhibited superior performance compared to the null model (χ² = 19.5, df = 2, p = 5.77*10⁻⁵). Subsequently, incorporating age as a second predictor further improved the model’s performance significantly when compared to the single-predictor model (χ² = 12.85, df = 2, p = .0016). Four-predictor model (APT, age, hemorrhage, and necrosis) significantly differed from the two-predictor model (APT and age) and led to improvement in classification performance (χ² = 22, df = 8, p = .005; accuracy = 0.74 [0.38, 0.92], mcAUC = 0.93). This final version of the model achieved sensitivity values ranging from 67% for the oligodendroglioma and astrocytoma groups to 87% for the glioblastoma group. The highest specificity score was observed in glioblastoma patients (93%), whereas the lowest specificity score was found in the astrocytoma class (79%) (the oligodendroglioma group had a value of 88%). The balanced accuracy magnitudes were 0.73 for astrocytoma, 0.77 for oligodendroglioma and 0.9 for glioblastoma (Figure 3(B)). The odds ratio for each predictor in the model can be found in Table 3.

Table 3.

Parameters of the best (four-predictors) model, classifying patients into three tumor types. Reference category was chosen to be the Astrocytoma group. Table provides the odds ratio, 95% confidence interval and corresponding p-values for each predictor in the model. OR—odds ratio, CI—confidence interval.

Characteristic	Glioblastoma	Oligodendroglioma	p-value	OR	95% CI	p-value
	OR	95% CI
Median APT	203	16.5, 249.7	0.03*	0.36	0.07, 1.83	0.2
Age	1.21	1.04, 1.41	0.013*	1.03	0.96, 1.11	0.4
Necrosis	7.04	0.70, 70.9	0.1	0.37	0.12, 1.19	0.095
Hemorrhage	0.01	0.00, 7.34	0.2	19.4	1.54, 243	0.022*

*statistical significance at the level of p < .05.

IDH1 status prediction

We further sought to predict the patient’s IDH1 class, that is, whether the patient had mutant or wild-type IDH1. For this end, several binary logistic regression models were developed. As were the cases with tumor grade and type classification, we started with APT median values being the only predictor in the model and found that this model was significantly different from the null model (χ² = 17.62, df = 1, p = 2.7*10⁻⁵). The inclusion of the age variable resulted in a substantial enhancement in the performance of the model (χ² = 12.56, df = 1, p = .0004). In total, these two variables provided a model with an accuracy of 0.79 [0.4, 0.97] and an area under the curve of 0.95; however, this model exhibited a low specificity rate of 30%. The likelihood ratio test indicated that further addition of variables to the model did not result in statistically significant differences compared to the model with only two variables. However, the inclusion of the predictor “necrosis” led to an increase in model accuracy up to 1 [0.66, 1], while there was minimal change in the area under the curve, which reached a value of 0.96 (χ² = 5.16, df = 3, p = .16). In Figure 4, we present the ROC curves for each of the aforementioned models. In addition, an increase in the median APT and age values by one unit led to the elevation of the odds ratio of membership in the wild-type group (see Table 4 for specific values).

Figure 4.

The plot displays areas under the curves for three binary logistic models (IDH1 group prediction) with different numbers of predictors, depicted in different colors. The dotted line corresponds to AUC = 0.5 or the random classifier’s performance.

Table 4.

Parameters of the best (two-predictors) model, classifying patients into two IDH1 groups. Table provides the odds ratio, 95% confidence interval and corresponding p-values for each predictor included in the model. Mutant type was chosen to be a reference level. OR—odds ratio, CI—confidence interval.

Characteristic	OR	95% CI	p-value
Median APT	32.5	4.78, 626	0.004**
Age	1.16	1.06, 1.33	0.009**

**statistical significance at the level of p < .01

Correlation between Ki-67 levels and APT values

We found a moderate positive correlation between Ki-67 and the 90th percentile APT values (r = 0.53 [0.26, 0.72], p = .003; Figure 5). Mean and median APT values also exhibited significant correlations with Ki-67 levels, although to a lesser extent than the 90th percentile values (r = 0.47 [0.18, 0.68], p = .004 for both variables, not shown).

Figure 5.

Figure displays the Spearman correlation between the 90th percentile APT and Ki-67 values. The grey area around the red line represents the 95% confidence interval. Individual data points are displayed as black dots.

Discussion

In this study, we explored the potential of using the APT technique to predict the grade and tumor type based on the WHO 2021 Classification of CNS Tumors for diffuse gliomas. We discovered that incorporating APT values into the models resulted in a substantial enhancement of their performance. Grade prediction metrics based on median APT values and patient age were the highest for grade 4, with grade 3 and grade 2 demonstrating moderate classification accuracy (all AUC values exceeded 0.75). These findings align well with the literature, which demonstrates the predictive potential of APT levels in distinguishing between low- and high-grade gliomas. ¹⁷ Furthermore, we expanded upon these findings by demonstrating the capability of APT to distinguish patients within the high-grade class, specifically those categorized as grade 3 and grade 4. It is crucial to consider this aspect, as there are significant disparities in both overall and median survival rates observed among patients diagnosed with grade 3 and grade 4 diffuse brain gliomas.^31–33 Similar results have been reported previously by Guo and colleagues in their study involving 62 patients, further validating our findings. ²⁵ However, the most notable contrast between their findings and ours was the existence of statistically significant differences between grade 2 and grade 3 patients in their study, a pattern that was absent in our data. This discrepancy could be attributed to the relatively small number of patients included in the grade 2 group within our study.

We were also able to identify the differences in APT levels among distinct tumor type groups. The most pronounced differences were found between the glioblastoma and oligodendroglioma groups, followed by astrocytoma and glioblastoma. In contrast, there were no statistically significant differences observed between the astrocytoma and oligodendroglioma groups. To our knowledge, these results have not been showcased in earlier studies and provide definitive clinical relevance.

IDH mutation and 1p19q codeletion statuses in patients with diffuse gliomas are well-established prognostic markers. ⁵ We discovered that patients with wild-type IDH1 had higher APT levels than those with mutant IDH1. Consistent with our current findings, several previously published studies have also shown a similar trend, indicating that elevated APT values can serve as a significant predictor of poor overall survival.^26,34 In the same vein, it has also been revealed that APT imaging exhibited superior performance over DKI in IDH mutation status prediction. ²⁴

Additionally, patients without codeletion were characterized by higher APT median values in comparison to patients with codeletion. However, contrary to what we have observed, Su and colleagues in their recent paper failed to identify differences between the groups in a sample of 113 patients with diffuse glioma. ³⁵ Further research is required to address this question, as the presence of 1p19q codeletion classifies the tumor as an oligodendroglioma, ⁵ which significantly improves prognosis for the patient.

Taken together, our findings clearly demonstrate the significant utility of the APT technique in predicting tumor grade and type, as well as identifying IDH mutation and 1p19q codeletion statuses, in patients with diffuse brain gliomas.

Hereafter, several multinomial logistic regression models were developed to predict patient tumor types and grades based on APT values complemented by additional parameters used as predictors. The model with the best classification performance for tumor grade, in addition to APT values, included the variable of patient age. This model demonstrated an accuracy of 0.71 and an area under the curve of 0.82, which is considered good classifier performance. The highest sensitivity values (88%) were obtained for Grade 4, while specificity was highest for Grade 3 (79%). Age is a known predictive factor for tumor malignancy and clearly impacts prognosis in patients with diffuse gliomas.^36–38 Therefore, entering the age of the patients in the APT-based prediction model looks logical, especially considering that this information is easy to obtain. However, we hypothesize that classification accuracy could be further improved by incorporating parameters not considered in this study.

In the case of tumor type classification, we settled on a model that includes four predictors. In addition to APT values, factors such as necrosis and the presence of hemorrhage were added. This allowed us to achieve an accuracy of 0.74 and an area under the curve of 0.93.

In general, we found that including additional parameters (specifically the patient’s age, proportion of tumor necrosis, and presence of hemorrhage) into the regression model significantly enhanced the accuracy of predicting type, grade, and IDH mutation status beyond just APT values.

We found positive associations between tumor proliferation molecular marker (Ki-67) levels and APT values. These results agree with previously published data. Several prior reports have also shown associations between APT values and Ki-67 levels.^23,39,40 These results support the notion that active tumor cell proliferation is linked to elevated concentrations of mobile proteins within the tumor, leading to an increased APT signal. Consequently, APT can indirectly reflect the proliferation index and subsequent malignancy of gliomas.

A previous study highlighted the APT technique as a viable alternative to contrast injection-dependent sequences such as DSC perfusion. ²⁰ However, our findings indicate that APT alone lacks the necessary precision to accurately predict tumor type, and the inclusion of necrosis proportion, which can only be identified through CE-T1WI, is needed. Exploring the potential utility of combining APT and DSC values in presurgical glioma definition represents a promising research direction warranting further investigation. Our study results reaffirm the significant potential of the APT technique for the noninvasive diagnosis of diffuse gliomas. Therefore, it should be routinely incorporated alongside morphological analysis. Another valuable application of APT lies in selecting optimal sites for presurgical stereotactic biopsies. It is well established that targeted biopsy of the most malignant region of a tumor enhances diagnostic accuracy, and this region can potentially be identified using APT maps. It is important to note that beyond its role in presurgical glioma evaluation, APT holds considerable utility in assessing treatment effects and overall prognosis,^26–28 although these aspects extend beyond the scope of our current study.

This study has several limitations. First, the sample size was small, consisting of only 42 patients, with only five cases presenting low-grade tumors. Second, our brain MRI protocol lacked a DSC perfusion sequence, thus impeding any comparison between the effectiveness of APT and rCBV values and hindering an assessment of their combined utility. Furthermore, there is a lack of follow-up data for the patients who participated in this study.

In conclusion, the amide proton transfer technique shows significant potential for noninvasive evaluation of diffuse gliomas. To the best of our knowledge, this study represents the initial endeavor to evaluate the effectiveness of APT in the preoperative assessment of gliomas using the WHO 2021 Classification of CNS Tumors. Furthermore, it is a pioneering attempt to integrate APT data with qualitative tumor characteristics and patient demographics to enhance diagnostic accuracy rates. However, given the pilot nature of the study, further studies with larger sample sizes and comprehensive follow-up are clearly necessary to strengthen these findings.

Appendix.

Abbreviations

ADC –

Apparent diffusion coefficient

APT –

Amide proton transfer

ASL –

Arterial spine labeling

CBF –

Cerebral blood flow

CBV –

Cerebral blood volume

CNS –

Central nervous system

DSC –

Dynamic susceptibility contrast

IDH –

Isocitrate dehydrogenase

MRI –

Magnetic resonance imaging

WHO –

World health organization

ORCID iD

Elena Filimonova https://orcid.org/0000-0002-6696-9071

Data Availability Statement

Raw data are available on https://openneuro.org, doi:10.18112/openneuro.ds004717.v1.0.0.