A Comparative Study of 2 Different Segmentation Methods of ADC Histogram for Differentiation Genetic Subtypes in Lower-Grade Diffuse Gliomas

Background To evaluate the diagnostic performance of apparent diffusion coefficient (ADC) histogram parameters for differentiating the genetic subtypes in lower-grade diffuse gliomas and explore which segmentation method (ROI-1, the entire tumor ROI; ROI2, the tumor ROI excluding cystic and necrotic portions) performs better. Materials and Methods We retrospectively evaluated 56 lower-grade diffuse gliomas and divided them into three categories: IDH-wild group (IDHwt, 16cases); IDH mutant with the intact 1p or 19q group (IDHmut/1p19q+, 18cases); and IDH mutant with the 1p/19q codeleted group (IDHmut/1p19q−, 22cases). Histogram parameters of ADC maps calculated with the two different ROI methods: ADCmean, min, max, mode, P5, P10, P25, P75, P90, P95, kurtosis, skewness, entropy, StDev, and inhomogenity were compared between these categories using the independent t test or Mann–Whitney U test. For statistically significant results, a receiver operating characteristic (ROC) curves were constructed, and the optimal cutoff value was determined by maximizing Youden's index. Area under the curve (AUC) results were compared using the method of Delong et al. Results The inhomogenity from the two different ROI methods for distinguishing IDHwt gliomas from IDHmut gliomas both showed the biggest AUC (0.788, 0.930), the optimal cutoff value was 0.229 (sensitivity, 81.3%; specificity, 75.0%) for the ROI-1 and 0.186 (sensitivity, 93.8%; specificity, 82.5%) for the ROI-2, and the AUC of the inhomogenity from the ROI-2 was significantly larger than that from another segmentation, but no significant differences were identified between the AUCs of other same parameters from the two different ROI methods. For the differentiaiton of IDHmut/1p19q− tumors and IDHmut/1p19q+ tumors, with the ROI-1, the ADCmode showed the biggest AUC (AUC: 0.784; sensitivity, 61.1%; specificity, 90.9%), with the ROI-2, and the skewness performed best (AUC, 0.821; sensitivity, 81.8%; specificity, 77.8%), but no significant differences were identified between the AUCs of the same parameters from the two different ROI methods. Conclusion ADC values analyzed by the histogram method could help to classify the genetic subtypes in lower-grade diffuse gliomas, no matter which ROI method was used. Extracting cystic and necrotic portions from the entire tumor lesions is preferable for evaluating the difference of the intratumoral heterogeneity and classifying IDH-wild tumors, but not significantly beneficial to predicting the 1p19q genotype in the lower-grade gliomas.


Introduction
Glioma is the most common neuroepithelial tumor in the brain which accounts for 80% of the malignant brain tumors. The severity of gliomas is further distinguished by malignant grades (I to IV) on the basis of the histopathological and clinical criteria [1]. The grade II and III gliomas are sometimes described as lower-grade gliomas, which present approximately one-third of all gliomas. Lower-grade gliomas form a biologically heterogeneous group of tumors. They are usually less aggressive tumors with a longer, indolent clinical course, but a subset of these gliomas will progress to glioblastoma (WHO grade IV gliomas) within months [2,3]. Histology alone is often insufficient to make accurate prognostic estimates, and tumors belonging to the same WHO grade may display different malignant behavior, depending on their molecular profile. The Cancer Genome Atlas (TCGA) Analysis Working Group grouped LGGs into three robust molecular classes on the basis of mutations in isocitrate dehydrogenases 1 and 2 (IDH1 and IDH2, hereafter collectively referred to as IDH) and codeletion of chromosomes 1p and 19q.
LGGs without IDH mutation are associated with the most aggressive clinical behavior and worst outcome, similar to that of glioblastomas (WHO grade IV).
LGGs with IDH mutation and 1p/19q codeletion are associated with the most favorable clinical outcome and possibly improved sensitivity to procarbazine, lomustine, and vincristine chemotherapy compared with noncodeleted neoplasms.
LGGs with IDH mutation and no 1p/19q codeletion are associated with an intermediate outcome, worse than those with 1p/19q codeletion, but far more favorable than IDHwt neoplasms [3]. This molecular classification has been integrated into the 2016 WHO classification of brain tumors [4].
Diffusion-weighted imaging (DWI) is a physiologic imaging modality that exploits the diffusion of water molecules to create contrast between tissues. Apparent diffusion coefficient (ADC) calculated from DWI is used as a quantitative parameter to assess the grade of restrictive diffusion and to provide information about tissue structure and cellularity [5,6]. Previous studies have demonstrated the ability of ADC to differentiate the IDH-wild gliomas from the mutant ones, the 1p19q codeleted gliomas from those noncodeleted ones [7][8][9]. However, previous studies were limited to using the mean value of ADCs based on regions of interest (ROIs) from a single representative slice of a lesion or region of interest from tumor volume, which may dilute or even mask the small but important differences between different disease entities. Additionally, they may not precisely depict the tumor status due to the intrinsic heterogeneous environment of tumors. Histogram analysis of the whole lesion may offer multiple parameters containing not only the quantitative accumulated ADC parameters, such as percentiles, minimal and maximal values, and mode but also the distribution parameters, such as the kurtosis, skewness, range, StDev, inhomogenity, and entrophy, thus providing more information about the tumor heterogeneity than the mean values [10,11]. In previous studies, two main ROI placement methods of the ADC histogram were used, including the tumor ROI excluding cystic and necrotic portions [12][13][14] and the entire tumor ROI containing cystic and necrotic areas [11,[15][16][17]. The theoretical basis of the former mentioned method is that necrotic and cystic-appearing areas may increase ADC values, which may be a confusing factor for differentiating the subtypes of gliomas based on ADC maps, but the latter method contained all compositions of the tumor, theoretically, it can better assess the heterogeneity of the tumor in its entirety. So, the purposes of this retrospective study were to evaluate whether the ADC values analyzed by the histogram method could help to classify IDH-wild tumors from IDH-mutated ones as well as IDHmut-NonCodel tumors from IDHmut-Codel ones in lowergrade diffuse gliomas and determine which segmentation method performs better.

Patients.
We searched the electronic hospital information system and picture archiving and communication system to identify patients from January 2016 to August 2019 who met the following inclusion criteria: (1) final histopathologic results were WHO grade II-III diffuse gliomas on the basis of the WHO classification for tumors of the central nervous system; (2) diffusion-weighted MRI, T2-weighted, and postcontrast T1-weighted anatomical scan performed at initial diagnosis and prior to any surgery; (3) and known IDH1 mutation and 1p/19q codeletion status. On the other hand, patients were excluded for the poor DWI images quality, which influence the consequent image analysis. The institutional review board at the First Affiliated Hospital of Chongqing Medical University approved this study. Thus, 56 consecutive patients were included in the final study cohort. The patients were divided into the following categories: IDH wild-group (IDH wt ), IDH mutant with the intact 1p or 19q group (IDH mut /1p19q + ), and IDH mutant with the 1p/19q codeleted group (IDH mut /1p19q − ).

ADC Histogram Measurement.
Each case was investigated using DWI (multishot echo-planar-imaging sequence with b values of 0 and 1000 s/mm 2 ) obtained with a 3.0 T MRI scanner (Signa HDxt, GE Medical System, WI). The ADC maps were digitally transferred from the picture archiving and communication system workstation to a personal computer and processed with an in-house software (Firevoxel, available at https://wp.nyu.edu/firevoxel/). For each case, the ROI was manually drawn by two independent radiologists with no knowledge of the final pathologic results. 2 different types of segmentation were completed: ROI-1, the entire tumor ROI containing all compositions; ROI-2, the entire tumor ROI excluding cystic and necrotic areas (cystic or necrotic portions met conditions: first, no enhancement with the contrast agent in the T1-weighted images and second, high intensity, like cerebrospinal fluid (CSF), in the T2-weighted images). According to Kang and Lue's methods, the tumor boundaries were defined with reference to the high-signal-intensity areas thought to represent the tumor tissue on the T2-weighted images [9,11]. The ROIs were placed carefully inside the mass to avoid regions influenced by the partial volume effect, and the position of the ROIs was verified using postcontrast T1-weighted images, T2weighted images, and T2-FLAIR imaging. Then, the  immunonegative tumors underwent multiple gene Sanger sequencing. The 1p/19q codeletion status was determined by fluorescence in situ hybridization-specific probes for the 1p36 and 19q13 loci.  [15,16]. The Shapiro-Wilk test was used to check whether the measurement data followed a normal distribution. Normally distributed continuous variables were compared using the independent t test, and nonnormally distributed continuous variables were compared using the Mann-Whitney U test between different molecular groups. For statistically significant results, receiver operating characteristic (ROC) curves were constructed to determine the optimal threshold for each histogram parameter to differentiate the molecular subtypes, and the optimal cutoff value was determined by maximizing Youden's index. Area under the curve (AUC) results were compared using the method of Delong et al. The results with p values of less than 0.05 were considered to be statistically significant.
For the IDH mut /1p19q + tumors, only the ADCmax ( p = 0:044) differed significantly between the two segmentation methods, and other histogram parameters did not show significant differences between the two segmentation methods (Table 3).

ROC Analysis and Comparation
. ADC histogram parameters from the two different ROI methods were evaluated for  (Tables 7 and 8). The inhomogenity from the two different ROI methods for distinguishing IDH wt gliomas from IDH mut gliomas both showed the biggest AUC (0.788, 0.930), and the optimal cutoff value was 0.229 (sensitivity, 81.3%; specificity, 75.0%) for the ROI-1 and 0.186 (sensitivity, 93.8%; specificity, 82.5%) for the ROI-2, respectively. In the pairwise comparison of ROC curves with the AUC, a major finding was that the AUC of the inhomogenity from the ROI-2 was significantly larger than that from the ROI-1, but no significant differences were identified between the AUCs of other same parameters from the two different ROI methods (Table 7).

Discussion
The results obtained from this study suggest that the ADC values analyzed by the histogram method can help to classify IDH-wild tumors from IDH-mutated tumors as well as IDHmut-Codel tumor from IDHmut-NonCodel tumors in lower-grade diffuse gliomas, no matter which ROI method is used. The StDev, inhomogenity, and range from the two different ROI methods were both larger in IDH-wild tumors compared with the IDH-mutated tumors, and the ADCmin and kurtosis in the IDH-wild group were smaller. Some different results appeared in the two different ROI methods, and one of these was that the P5ADC from the ROI-2 in the IDH-wild tumors was lower than the mutated ones, while P5ADC from ROI-1 showed no significant difference in these two types of tumors. Another difference was that the 9 BioMed Research International P75ADC, P90ADC, P95ADC, and ADCmax from ROI-1 were larger in IDH-wild tumors, while the same parameters from the ROI-2 showed no significant differences in the two sorts of tumor. For identifying the 1p19q-Codel ones in the IDH-mutated tumors, we found that the ADCmean, P5ADC, P10ADC, P25ADC, P50ADC, P75ADC, P90ADC, and ADCmode from the two different ROI methods were both significantly lower in the 1p19q -Codel tumors; meanwhile, the skewness was greater in this group. Compared with ROI-1, the second segmentation gave us more parameters valuable for the differentiation, such as the P95ADC and kurtosis, the P95ADC was significantly lower in the 1p19q-Codel tumors, and the kurtosis was lager in the 1p19q-Codel ones.
It is widely recognized that the pathological heterogeneity may manifest as radiologic heterogeneity on ADC maps [5], and differences in ADCs are mainly attributed to the tumor cellularity but also to the presence of necrosis or cysts [6,11,18]. Consequently, the ADCs within a given tumor can vary widely between different regions of that tumor [10]. The voxels with low ADC value are reportedly well correlated with highly cellular components within the tumor, which reflects tumor proliferative rate and aggressiveness [5], whereas the higher frequency of voxels with high ADC values reflect cystic, necrotic, or myxoid components. In other words, the larger intratumoral heterogeneity is the wider ADC values distribute [15]. Our results demonstrated that the IDH wild-type gliomas showed a lower ADCmin, likely representing higher cellularity and aggressiveness, consistent with some previous studies [7,9,19], as well as appeared larger higher end values, inhomogenity, and StDev, likely representing more common cystic and necrotic portions, which might be associated with higher grade features and larger intratumoral heterogenitiy [11,20]; meanwhile, in the present study, we found that no matter which segmentation was used, the parameters reflecting the intratumoral heterogenitiy, such as the inhomogenity and StDev, performed better than the conventional cumulative ADC values, so it is valuable and important to evaluate the intratumoral heterogenitiy using quantitative parameters. It was interesting that the entire tumor ROI provided some additional statistically significant parameters than the other segmentation, such as the higher end ADC values. This result suggests that, to a certain degree, the necrotic and cystic components may facilitate the discrimination between IDH-mutated tumors and IDH-wild gliomas, but their performance is weaker than the inhomogenity. The inhomogenity calculated from the ROI-2 performed better than that from the ROI-1, although the diagnostic performance of the other parameters obtained from the two methods showed no statistically significant differences, which reveals that extracting cystic and necrotic portions from the entire tumor lesions is better for evaluating the difference of the intratumoral heterogeneity and more helpful to classify IDH-wild tumors in the lower-grade gliomas.
For the discrimination between the IDHmut-NonCodel tumors and the IDHmut-Codel tumors, an important finding in our study was the increase of restrictive diffusion in the  [14,21]. There are several possible explanations for this result. One of them is the presence of calcification, which is common in IDHmut-Codel tumors [22], that may limit water content as well as hinder water movement [20], and another is that the IDHmut-Codel tumors are often highly cellular lesions with closely packed, relatively small cells in central regions and prominent secondary structure formation, which may also delay the passage of small molecules [23]. When applying the ROI-1, the ADCmode performed best at predicting the 1p19q genotype in this study, which was lower in the IDHmut-Codel tumors. ADCmode means the value that appears most frequently in a set of ADC values, and this result may also be due to the larger diffusion restriction in the IDHmut-Codel tumors. With the second segmentation, the skewness showed the biggest AUC, which was larger in IDHmut-Codel tumors. The skewness reflects the asymmetry of the ADC distribution, more pixels have lower ADC values and lie to the left of the mean of the histogram, and the skewness is more positive. Therefore, this result indicates that the IDHmut-Codel tumors contain more voxels with ADC values below the mean of the histogram, which may also be associated with the increase of restrictive diffusion, but the performance of the same parameters from the two different method showed no statistically significant differences. As for the comparison of histogram parameters calculated from the two different segmentations, we found that almost all ADC histogram parameters calculated from the whole tumor ROIs tended to be larger than those from the other segmentation. Statistically, significant differences were found in the higher end values of cumulative ADC histograms and some histogram distribution characteristics from these two different methods in IDH-wild gliomas and IDHmut-Codel tumors, such as P75ADC, P90ADC, P95ADC, ADCmax, StDev, and inhomogenity. This result may be associated with the following reasons: the first one is that the IDH wild-type gliomas show larger intratumoral heterogeneity due to its increased cell proliferation and necrosis [7,20], so the entire tumor ROI contained areas of necrosis and systs will lead to more higher ADC values, larger StDev, and inhomogenity. As for the codeleted tumors, the signal intensity on MR images is more heterogeneous than the noncodeleted ones [20,24]; as a result, the StDev and inhomogenity obtained from the entire tumor ROI were lager. Conversely, the similar results of the two ROI segmentations in noncodeleted tumors can be explained by a relative homogeneity in these gliomas [20], so extracting cystic and necrotic portions from the ROIs does not cause any obvious differences.