Abstract

   We aimed to predict CD44 expression and assess its prognostic
   significance in patients with high-grade gliomas (HGG) using
   non-invasive radiomics models based on machine learning. Enhanced
   magnetic resonance imaging, along with the corresponding gene
   expression and clinicopathological data, was downloaded from online
   database. Kaplan–Meier survival curves, univariate and multivariate COX
   analyses, and time-dependent receiver operating characteristic were
   used to assess the prognostic value of CD44. Following the screening of
   radiomic features using repeat least absolute shrinkage and selection
   operator, two radiomics models were constructed utilizing logistic
   regression and support vector machine for validation purposes. The
   results indicated that CD44 protein levels were higher in HGG compared
   to normal brain tissues, and CD44 expression emerged as an independent
   biomarker of diminished overall survival (OS) in patients with HGG.
   Moreover, two predictive models based on seven radiomic features were
   built to predict CD44 expression levels in HGG, achieving areas under
   the curves (AUC) of 0.809 and 0.806, respectively. Calibration and
   decision curve analysis validated the fitness of the models. Notably,
   patients with high radiomic scores presented worse OS (p < 0.001). In
   summary, our results indicated that the radiomics models effectively
   differentiate CD44 expression level and OS in patients with HGG.

Supplementary Information

   The online version contains supplementary material available at
   10.1038/s41598-025-90128-7.

   Keywords: Radiomics, CD44, High-grade glioma, Overall survival, Machine
   learning

   Subject terms: Immunology, Biomarkers, Oncology, Cancer, CNS cancer

Introduction

   Glioma is the most common primary malignancy occurring in the central
   nervous system (CNS)^[32]1. Despite advancements in surgery,
   chemotherapy, radiotherapy, and targeted therapy for glioma in recent
   decades, it remains a highly lethal disease due to limited treatment
   efficacy. Gliomas are categorized into grade I to IV based on the
   histologic and molecular features^[33]2. Generally, grade I and II
   gliomas are categorized as low-grade glioma (LGG) and have a relatively
   favorable prognosis. In contrast, grade III and IV gliomas are
   classified as high-grade gliomas (HGG), which are highly malignant,
   aggressive and associated with a miserable prognosis. According to the
   fifth edition of the WHO Classification of Tumors of CNS, the
   mutational status of IDH is one of the most crucial biomarkers for
   gliomas, as it helps distinguish GBM from both astrocytoma and
   oligodendroglioma^[34]3. Currently, the prognostic indicators of glioma
   include clinicopathologic features, isocitrate dehydrogenase (IDH)
   status, O^6-methylguanine DNA methyltransferase promoter (MGMT) status,
   Chromosome (Chr) 1p19q codeletion and radiomics data of computer
   tomography (CT) and magnetic resonance imaging (MRI). However, these
   indicators are insufficient for precision medicine due to the
   heterogeneity of gliomas. Consequently, there is an urgent need to
   investigate novel predictors to improve prognostic stratification and
   personalize medical decision-making.

   The protein CD44 antigen (CD44) is a glycoprotein located on the cell
   surface, which functions as a receptor for extracellular matrix
   components such as hyaluronan, collagen, and matrix
   metalloproteinases^[35]4,[36]5. It participates in various biological
   functions, including cell adhesion, migration, lymphocyte activation,
   cell-cell interactions, hematopoiesis, and inflammation^[37]6–[38]9. A
   prior study discovered that high CD44 expression is linked to
   diminished survival rates in patients with grade II and III glioma,
   indicating its potential as a prognostic marker^[39]10. Several
   compounds targeting CD44 yielded promising antitumor
   effects^[40]11–[41]14.

   For instance, a phase II clinical study highlighted the efficacy and
   safety of Angstrom6 in ovarian cancer by modulating CD44-mediated
   signaling^[42]11. Nevertheless, detection of CD44 through blood is
   inconvenient, expensive, and subjected to spatio-temporal heterogeneity
   with tumor. Moreover, detection of CD44 by tissue sample is invasive,
   costly, expensive, and influenced by factors such as antibody
   specificity and operator variability. Consequently, an efficient,
   non-invasive, and repeatable method to predict CD44 expression in
   patients with HGG is urgently needed.

   Given its low cost, non-invasiveness, and effectiveness, MRI is
   routinely implemented in the patients with glioma. Radiomics offers a
   method for quantitatively, dynamically, and non-invasively reporting
   tumor characteristics based on advanced imaging features by analyzing a
   vast array of radiographic characteristics^[43]15. Machine learning
   (ML) has been extensively applied in clinical diagnosis and precision
   medicine, especially in the fields of radiology and oncology.
   Increasing studies have indicated the potential of non-invasive
   evaluation via radiomics for early diagnosis and molecular typing of
   HGG, as well as for predicting the tumor microenvironment (TME) and
   heterogeneity^[44]16–[45]18. However, there have been no published
   studies on predicting CD44 expression in patients with HGG using
   radiomics.

   In the present study, we aimed to build a non-invasive radiomics model
   by ML to predict CD44 expression in patients with HGG. This was
   achieved by utilizing data from The Cancer Genome Atlas (TCGA) and The
   Cancer Imaging Archive (TCIA) databases. Additionally, we explored the
   potential relationship between CD44 and the tumor microenvironment
   (TME), which could serve as a valuable predictive tool to enhance
   clinical decision-making and precision medicine.

Materials and methods

Data collection

   MRI images, along with corresponding gene expression data and
   clinicopathological data of patients with HGG, were downloaded from the
   TCIA ([46]https://www.cancerimagingarchive.net/) and TCGA
   ([47]https://portal.gdc.cancer.gov/) databases. The inclusion criteria
   for this study were: (1) newly diagnosed HGG; (2) pathologically
   confirmed grade III or IV gliomas; (3) availability of axial T1WI
   contrast-enhanced (T1WI + C) images of diagnostic quality. The
   exclusion criteria were: (1) absence of gene expression data, clinical
   information, and survival data; (2) survival time of less than 30 days;
   and (3) low image quality. The schematic drawing of the criteria was
   listed in Supplementary Figure [48]S1.

   For each patient, we collected the following variables: age (< 60 or
   ≥ 60 years), gender (female or male), grade (III or IV), IDH status
   (wild-type or mutated), Chr 1p19q codeletion status (codeletion or
   non-codeletion), the methylation status of MGMT promoter
   (unmethylated/unknown or methylated), chemotherapy (no or yes),
   radiotherapy (no or yes), expression level of CD44, survival time, and
   survival outcome (alive or death). The requirement for ethical approval
   was waived.

Expression analysis of CD44

   The RNAseq data (TPM) from TCGA and Genotype-Tissue Expression (GTEx)
   processed by the TOIL project^[49]19, were accessed from UCSC Xena
   ([50]https://xenabrowser.net/datapages/). All the patients were divided
   into CD44^high and CD44^low groups based on the cutoff expression level
   of CD44 determined by the survminer R package. Differential CD44
   expression analysis was performed between HGG samples and normal brain
   samples (NBTs), following the log2 transformation of the RNAseq data.

Immunohistochemistry (IHC)

   We obtained a total of 4 HGG tissue samples and 4 NBTs from patients
   with epilepsy and brain injury in the Department of Pathology and
   Neurosurgery, Shengjing Hospital of China Medical University. After
   fixation in 10% formalin, the samples were paraffin-embedded and
   sectioned. Subsequently, IHC staining with an anti-CD44 antibody
   (Proteintech, # 60224-1-Ig, dilution 1:200) was performed to assess
   CD44 expression. Informed consent was obtained from all patients, and
   the study was approved by the Ethics Committee of Shengjing Hospital
   (No. 2019PS105K).

Survival and enrichment analysis of CD44

   OS was assessed and plotted using Kaplan–Meier survival curves. Cox
   univariate and multivariate regression analyses were employed to
   evaluate the influence of various variables on OS. Moreover, we
   conducted subgroup analyses and interaction tests to evaluate the
   effects of CD44. The time-dependent receiver operating characteristic
   (ROC) curves were performed to estimate the accuracy and discrimination
   of CD44 in predicting OS at 12, 24, and 36 months of follow-up. To
   determine the correlations between CD44 and clinical features, we
   calculated Spearman rank coefficients.

   We applied CIBERSORTx ([51]https://cibersortx.stanford.edu/) to
   deconvolute of the expression matrix of HGG samples, thereby obtaining
   the distribution of 20 distinct immune cell types in HGG tissues of
   each sample^[52]20. We conducted Wilcoxon test to assess the immune
   cell abundances between the CD44^high and CD44^low groups by the R
   package limma^[53]21. To identify the different functions and relevant
   pathways between the CD44^high and CD44^low groups, we conducted gene
   ontology (GO) and the Kyoto Encyclopedia of Genes and Genomes (KEGG)
   pathway analysis on differentially expressed genes (DEGs) by the R
   package clusterProfiler with q value < 0.05^[54]22.

Development of a radiomics models to predict CD44 expression level in
patients with HGG

   Totally, we obtained 82 samples by intersecting the TCGA bioinformatics
   data with TCIA image data. The imaging parameters were as follows:
   Slice Thickness: 0.9–6.0 mm; Pixel_Spacings: 0.4–1.1 mm; Spacing
   Between Slices: 1.0–7.5 mm; RepetitionTime: 6.6-3252.8 ms; Echo Time:
   2.75-20.0 ms. We applied N4 bias field correction to the MR images to
   reduce signal inhomogeneity caused by magnetic field variations, which
   can lead to regions appearing too bright or dark, or inconsistencies in
   brightness within the same tissue. Images were resampled to achieve
   isotropy, minimizing variations due to differences in scanning
   equipment and protocols, as well as lesion size. Image normalization
   was also implemented to reduce noise and standardize intensity, thereby
   minimizing differences in image signal intensity across different
   machines. Patients were divided into CD44^high and CD44^low groups
   based on the cutoff expression level of CD44 determined by the
   survminer R package.

   Two experienced physicians manually outlined entire tumor areas in the
   MRI enhanced T1-weighted images (T1WI)^[55]23 using three-dimensional
   slicer version 4.10.2. The Volume of Interest (VOI) delineation was
   performed by two radiologists who were blinded to the clinical data.
   One radiologist completed the VOI delineation for all cases, while the
   second radiologist randomly selected 20 samples for independent
   delineation. A total of 107 radiomic features were extracted by
   PyRadiomics and subsequently normalized. The intraclass correlation
   coefficient (ICC) was utilized to evaluate the reliability of the
   radiomic features between the two physicians, and features with
   ICC ≥ 0.75 were chosen for subsequent analyses.

   Repeat least absolute shrinkage and selection operator (LASSO) was
   carried out to further screen the radiomic features by R package
   glmnet^[56]24. Features that occurred more than 600 times during the
   1000 screening were included in the final model. After feature
   extraction, Z-score normalization was utilized.

   Logistic Regression (LR) and Support vector machine (SVM) were used to
   predict the expression of CD44. The diagnosis discrimination of
   radiomics models was assessed by plotting the ROC and precision recall
   (PR) curves. The calibration curve was generated to evaluate the
   goodness of fit of radiomics models. Brier score was used to determine
   the overall performance of radiomics models. The clinical benefit was
   determined by a decision curve analysis (DCA). To identify the
   predicted CD44 levels by the radiomic models, we calculated radiomic
   score (RS). The Wilcoxon test was utilized to assess the difference of
   RS between CD44^high and CD44^low groups. To determine the correlations
   among RS, CD44 and genes of EMT, we performed Spearman correlation
   analysis.

   We obtained 170 samples by intersecting TCGA clinical data and TCIA
   imaging data. The 170 patients were divided into dichotomous variable
   (i.e., RS^high and RS^low groups) based on the cutoff value of RS
   derived from SVM model determined by R package survminer. Log-rank
   tests were used to survival analysis.

Statistics analysis

   The quantitative data were described as mean ± standard deviation. The
   categorical data were described as frequencies and proportions, and
   then compared using χ^2 test at baseline. Statistical significance was
   considered as a two-sided p-value < 0.05. The T1WI + C-clinical
   prediction model was visualized through a nomogram by the rms R
   package. All statistical analyses were performed by R software version
   4.1.1.

Results

Patient characteristics

   The flow-chart of this study was presented in Fig. [57]1. A cohort of
   310 HGG patients from the TCGA databases were included in survival
   analyses. Patients were separated into CD44^high (n = 155) and CD44^low
   (n = 155) groups based on the median expression of CD44 (5.043). The
   baseline characteristics of the patients were detailed in Supplemental
   Table 1.

Fig. 1.

   [58]Fig. 1
   [59]Open in a new tab

   The brief flowchart of the study process.

Clinical significance of CD44

   To elucidate the clinical relevance of CD44, we initially investigated
   its expression in HGG samples and NBTs. The results showed that CD44
   expression was markedly elevated in HGG samples (p < 0.001,
   Fig. [60]2A). Furthermore, we conducted immunohistochemistry (IHC)
   experiments on HGG patient samples and NBTs, confirming that CD44
   protein levels were elevated in HGG compared to NBTs (Fig. [61]2B). The
   median OS in the CD44^low group was 44.6 months, compared to just 16.2
   months in the CD44^high group. Patients in the CD44^low group exhibited
   better OS than those in the CD44^high group (p < 0.001, Fig. [62]2C).
   Additionally, we analyzed the impact of various clinical
   characteristics to the OS of HGG patients. The results indicated that
   low age (< 60), histologic grade, chemotherapy, IDH mutant, Chr 1p19q
   codeletion, and methylated MGMT promoter were associated with longer OS
   (Supplementary Figure [63]S2). The time-dependent ROC curve indicated
   that the area under the curve (AUC) of the CD44 expression was 0.719
   for 1-year OS prediction (Fig. [64]2D).

Fig. 2.

   [65]Fig. 2
   [66]Open in a new tab

   Expression of CD44 elevated and predicted poor OS in HGG. (A) CD44 was
   high expression in HGG tissues. (B) IHC analysis revealed that CD44 was
   upregulated in HGG compared to that in normal brain tissues. Images
   were captured at × 100. (C) Patients in the CD44^low group had better
   OS. (D) The ROC curves of CD44 in predicting OS at 12, 24, and 36
   months of follow-up. (E) Cox univariate and multivariate regression
   analyses of the effects of clinicopathological variables on OS. (F)
   Subgroup analyses and interaction tests of the effects of CD44
   expression level on OS. (G) The correlations between CD44 expression
   and clinical characteristics of HGG. CI, confidence interval; HGG,
   high-grade glioma; HR, hazard ratio; IDH, isocitrate dehydrogenase;
   IHC, immunohistochemistry; MGMT,
   O^6-methylguanine-DNA-methyltransferase; OS, overall survival; ROC,
   time-dependent receiver operating characteristic. *, p < 0.05; **,
   p < 0.01; ***, p < 0.001.

   Further, the univariate Cox analysis revealed that compared to low CD44
   expression, high CD44 expression was significantly correlated with
   worse OS (HR = 3.025, 95% CI: 2.203–4.154, p < 0.001). Expectedly, low
   age (< 60), histologic grade, chemotherapy, IDH mutant, Chr 1p19q
   codeletion, and methylated MGMT promoter presented similar trend. In
   multivariate Cox analysis, high CD44 expression emerged as an
   independent factor of shorter OS compared to low CD44 expression
   (HR = 1.718, 95% CI: 1.189–2.484, p = 0.004, Fig. [67]2E). Moreover,
   low histologic grade, chemotherapy, and IDH mutant were identified as
   independent factors affecting OS in HGG patients. Subgroup analyses
   revealed that most subgroups with high CD44 expression exhibited
   unfavorable OS. In exact, significant interaction effects of CD44
   expression on the OS were seen in chemotherapy and radiotherapy
   subgroups (Fig. [68]2F). The heatmap illustrating the correlation
   between CD44 expression and clinical characteristics revealed that CD44
   expression was negatively and significantly related to IDH status, Chr
   1p19q codeletion, and MGMT promoter status (all p < 0.001,
   Fig. [69]2G).

Immune infiltration and enrichment analysis

   Subsequently, we delved into the distinctions in the tumor immune
   microenvironment between the CD44^high and CD44^low groups. As shown in
   Fig. [70]3A, the violin plots revealed a higher infiltration of M2
   macrophages, monocytes, and CD4 memory resting T cells in the CD44^high
   group compared with the CD44^low group (p < 0.001).

Fig. 3.

   Fig. 3
   [71]Open in a new tab

   The immune and functional analyses of DEGs between the CD44^high and
   CD44^low groups. (A) The correlations of CD44 expression and immune
   cell infiltration in HGG. (B) The GO analyses of DEGs between the
   CD44^high and CD44^low groups. (C) The KEGG analyses of DEGs between
   the CD44^high and CD44^low groups. BP, biological process; CC, cell
   component; DEGs, differentially expressed genes; GO, gene ontology;
   HGG, high-grade glioma; KEGG, the Kyoto Encyclopedia of Genes and
   Genomes; MF, molecular function.

   To further understand the biological functions and mechanisms of CD44
   in HGG, we performed GO and KEGG analyses of DEGs (listed in
   Supplemental Table 2) between the CD44^high and CD44^low groups.
   Synapse organization, focal adhesion, and transcription coregulator
   activity was significantly enriched in GO analysis (p < 0.05,
   Fig. [72]3B). KEGG pathway enrichment analysis unveiled several
   enriched pathways, including lysosome, Epstein-Barr virus infection,
   adherens junction, and Human immunodeficiency virus 1 infection
   (p < 0.05, Fig. [73]3C).

Construction and evaluation of radiomics models

   Totally, 82 patients from the intersection of TCGA and TCIA datasets
   were enrolled for radiomics analysis. Patients were separated into
   CD44^high and CD44^low groups based on the median expression of CD44
   (5.043). The ICCs of the features extracted by PyRadiomics were listed
   in Supplemental Table 3. The median of ICC was 0.938, and 93 features
   (86.9%) had an ICC of ≥ 0.75, which were enrolled for further feature
   selection. After repeat LASSO, seven features with frequencies > 600
   remained in the model, including shape_sphericity, shape_sphericity,
   gray level size zone matrix (glszm)_gray level non uniformity
   normalized, and gray level run length matrix (glrlm)_long run high gray
   level emphasis, among others (Fig. [74]4).

Fig. 4.

   [75]Fig. 4
   [76]Open in a new tab

   Selection of radiomic features by LASSO. (A) Plot of the LASSO
   coefficient profiles of the selected features. The vertical line
   presented the optimal λ value. (B) Selection of tuning parameter (λ) in
   the LASSO model. The vertical line presented the optimal λ value. (C)
   Visualization the top 20 radiomic features with the most counts. The
   dashed line presented 600 counts. LASSO, least absolute shrinkage and
   selection operator.

   Subsequently, we constructed a radiomics model based on LR. The
   importance of seven features were showed in Fig. [77]5A. The radiomics
   score (RS) was calculated using the formula: RS =
   -0.01933 + 0.880206×shape_sphericity + 0.072922×glszm_gray level non
   uniformity normalized + 0.165961 × glrlm_long run high gray level
   emphasis-1.05154 × neighborhood gray tone difference matrix
   (ngtdm)_coarseness-0.5184 × glszm_large area low gray level
   emphasis-0.44085 × shape_surface area + 0.483907×first order mean. The
   model demonstrated an accuracy of 0.768, with sensitivity of 0.829,
   specificity of 0.707, positive predictive value (PPV) of 0.739, and
   negative predictive value (NPV) of 0.806. The predictive performance of
   the model was satisfactory, achieving an AUC of 0.806 (Fig. [78]5B).
   Moreover, the AUC in the 10-fold cross-validation and PR curve reached
   0.751 and 0.772 (Fig. [79]5C,D).

Fig. 5.

   [80]Fig. 5
   [81]Open in a new tab

   Construction and assessment of the radiomics model for predicting CD44
   expression based on LR. (A) Seven radiomic features and the
   corresponding importance in the radiomics model based on LR. (B) ROC
   curve. (C) ROC curve of 10-fold cross-validation. (D) PR curve. (E)
   Calibration curve. (F) Plot of DCA. (G) Predicted probabilities of RS
   in CD44^high and CD44^low groups. AUC, the area under the curve; DCA,
   decision curve analysis; Glrlm, gray level run length matrix; Glszm,
   gray level size zone matrix; HL, Hosmer-Lemeshow test; LR, logistic
   regression; Ngtdm, neighborhood gray tone difference matrix; PR,
   precision recall; ROC, receiver operating characteristic. ****,
   p < 0.0001.

   The calibration curve and Hosmer-Lemeshow test indicated a favorable
   consistency between the model-predicted probabilities of CD44 high
   expression and the actual values (p = 0.161, Fig. [82]5E). The DCA
   curves suggested our radiomics model provided higher net benefit than
   all treatments and none treatment within a threshold probability was
   0.1–0.7 (Fig. [83]5F). Notably, the RS was significantly higher in the
   CD44^high group than in the CD44^low group (p < 0.001, Fig. [84]5G).

   To further investigate the optimal radiomics model for predicting CD44
   expression in HGG, we developed a radiomics model utilizing SVM. The
   importance of seven features were showed in Fig. [85]6A. The SVM-based
   radiomics model achieved an accuracy of 0.780, with sensitivity at
   0.854, specificity at 0.707, PPV at 0.745, and NPV at 0.829. Similarly,
   the predictive performance of the model was good with an AUC of 0.809
   (Fig. [86]6B). In addition, the AUC of 10-fold cross-validation and the
   PR curve reached 0.778 and 0.771. (Fig. [87]6C,D). The calibration
   curve and Hosmer-Lemeshow test indicated a favorable consistency
   between the predictive CD44 expression and the true value (p = 0.194,
   Fig. [88]6E). The DCA curves suggested that our SVM-based radiomics
   model provided a higher net benefit than all treatments and none
   treatment within a threshold probability was 0.1–0.8 (Fig. [89]6F). The
   RS was significantly higher in the CD44^high group than in the CD44^low
   group (p < 0.001, Fig. [90]6G). Although the AUC for the SVM model was
   higher than that for LR model, no significant difference was found
   between the two models by Delong test. The RS of SVM model was chosen
   for further analyses due to its relatively higher accuracy.

Fig. 6.

   [91]Fig. 6
   [92]Open in a new tab

   Construction and assessment of the radiomics model for predicting CD44
   expression based on SVM. (A) Seven radiomic features and the
   corresponding importance in the radiomics model based on SVM. (B) ROC
   curve. (C) ROC curve of 10-fold cross-validation. (D) PR curve. (E)
   Calibration curve. (F) Plot of DCA. (G) Predicted probabilities of RS
   in CD44^high and CD44^low groups. (H) Spearman correlation analysis of
   RS and EMT molecules. (I) Survival analysis of RS for OS. (J) Clinical
   application of radiomics model in HGG. AUC, the area under the curve;
   DCA, decision curve analysis; Glrlm, gray level run length matrix;
   Glszm, gray level size zone matrix; HL, Hosmer-Lemeshow test; Ngtdm,
   neighborhood gray tone difference matrix; OS, overall survival; PR,
   precision recall; ROC, receiver operating characteristic; RS, Radiomic
   score; SVM, support vector machine; ***, p < 0.001; ****, p < 0.0001.

   In addition to CD44, Spearman correlation analysis illustrated that RS
   was also significantly correlated to some molecules involved in EMT
   (p < 0.05), including serpin H1 (SERPINH1), insulin like growth factor
   binding protein 2 (IGFBP2), galectin-1 (LGALS1). (Fig. [93]6H). To
   investigate the prognostic implications of the radiomics model, we
   obtained 170 patients’ data with HGG from TCGA database for subsequent
   survival analysis. All the patients were divided into RS^high (n = 85)
   and RS^low (n = 85) groups based on the cutoff value of 0.532. The
   baseline clinicopathological characteristics of the patients were
   presented in Supplemental Table 4. As shown in Fig. [94]6I, the median
   OS was significantly elevated in RS^low group (21.6 months) compared to
   RS^high group (14.9 months), indicating that patients in RS^high group
   had poorer OS than those in RS^low group (p < 0.001).

   To enhance the clinical applicability, we developed a predictive
   nomogram incorporating radiomic score, age, gender, chemotherapy,
   radiotherapy, the methylation status of MGMT promoter, and Chr 1p19q
   codeletion status. The clinical application nomogram score sheet
   displayed the clinical predictive scheme at 12, 36, and 60 months
   (Fig. [95]6J).

Discussion

   The present study aimed to elucidate the prognostic value of CD44 in
   HGG, and to develop radiomics models by ML to predict CD44 expression
   non-invasively. Our study corroborated that CD44 expression level was
   an independent predictor of OS, and high-expressed CD44 was associated
   with poor OS among HGG patients. In addition, we constructed two
   favorable radiomics models to predict CD44 expression level in HGG
   patients, achieving AUC values of 0.809 and 0.806, respectively.
   Furthermore, HGG patients with higher RS predicted relatively shorter
   OS.

   CD44, a cell-surface receptor with multifaceted roles in cellular
   functions, has been implicated as a risk factor in glioma, as
   demonstrated in previous studies^[96]25,[97]26. For instance, a
   meta-analysis revealed that elevated CD44 expression adversely impacts
   OS in patients with glioma^[98]25, which aligns with our findings.
   Accumulating evidence suggests that CD44 is responsible for growth and
   metastasis in HGG^[99]27,[100]28. Interestingly, our analysis found
   that the most significant GO enrichment terms associated with CD44
   include synapse organization, focal adhesion, cell-substrate junction,
   and neuron to neuron synapse. In addition, CD44 was recently reported
   to participate in TME^[101]29, immunotherapy^[102]30, and
   stemness^[103]31 in HGG. Notably, our research also discovered a
   propensity for M2 macrophages infiltration in the CD44 high-expressed
   HGG, indicating a potential role of CD44 in driving M2 macrophage
   polarization and remodeling TME^[104]29.

   Numerous molecules, including CD44, contribute to HGG progression in
   distinct ways and at different stages, contributing to intratumoral and
   intertumoral heterogeneity^[105]32. However, assessing this
   heterogeneity pathologically is invasive, subject to sampling
   variability, and impractical in clinical practice for each sample.
   Thus, it is essential to construct an efficient approach to quantify
   the heterogeneity. MRI is routinely utilized for HGG diagnosis and
   follow-up in clinical practice, offering the capability to
   non-invasively assess the entire lesion and dynamic monitoring it.
   Several evidences have inspired us to explore the potential of
   MRI-based radiomics for diagnosis and classification of
   HGG^[106]33,[107]34.

   In the current research, we identified CD44 as an independent risk
   factor for OS by multivariate COX regression analysis. To further
   predict CD44 expression in HGG, we applied repeat LASSO to screen the
   radiomic features of MRI images, and constructed two radiomics models
   based on LR and SVM methods, respectively. Interestingly, both models
   exhibited satisfactory predictive power of CD44 expression. As
   precision medicine continues to advance, an increasing number of
   molecular targets have been identified. Notably, accurate prediction of
   these targets has become critical for diagnosis and therapy, forming
   the foundation of precision oncology applications. Consequently, many
   investigators are committed to build radiomics models to predict
   molecule status intratumor precisely and
   generalizability^[108]35,[109]36. Thus, radiomics models are emerging
   as a popular research direction for heterogeneity measure and biomarker
   prediction.

   Routine MRI diagnosis relies on morphological features of the lesions,
   which is lack of objective, quantitative information. Radiomics
   provides high-dimensional quantitative information that might have been
   overlooked or unnoticed by the naked eye, thereby aiding diagnosis
   based on imaging features. It is often referred to as “digital biopsy”.
   Recently, extensive studies have confirmed the importance of radiomics
   in predicting biomarkers in glioma. For instance, Sha et al.^[110]37
   constructed radiomics models to predict the IDH mutation combined with
   MGMT methylated molecular subtype in glioma, achieving an impressive
   AUC of 0.9. Another study built an MRI-based radiomic signature that
   had promising ability to assess the promoter mutation status of TERT in
   glioblastoma patients^[111]38. Similarly, a recent study built two
   radiomics models based on MRI to predict CD44 expression in
   LGG^[112]39. In this research, we applied MRI radiomic features to
   predict CD44 expression in HGG, and our prediction models achieved AUCs
   of 0.806 and 0.809 in LR and SVM models, respectively. These results
   provide an efficient, cost-effective, repeatable, non-invasive method
   to predict CD44 expression in HGG.

   Following the screening of 107 radiomic features, seven radiomic
   features were finally retained to build our predictive model, including
   shape_sphericity, glszm_gray level non uniformity normalized,
   glrlm_long run high gray level emphasis, ngtdm_coarseness, glszm_large
   area low gray level emphasis, shape_surface area, and first order mean.
   Glszm, a texture feature, represents the number of connected voxels
   with the same gray level intensity and arrangement pattern in an
   image^[113]40. A higher value indicates greater heterogeneity within
   regions of interest. Several studies confirmed that glszm-derived
   features are common and important radiomic factors for the research in
   glioma^[114]39,[115]41. Wang et al. built a predictive model for CD44
   expression in LGG, in which a feature from glszm category showed an
   odds ratio of 2.634 (p = 0.008)^[116]40. Consistent with their
   findings, our study also highlighted the crucial role of the glszm
   category in both models. In the SVM model, glszm_gray level non
   uniformity normalized emerged as the most important feature, while in
   the LR model, glszm_large area low gray level emphasis ranked as the
   third most important feature. The above results underscore that the
   glszm radiomic features played an important role in predicting the CD44
   expression level in HGG.

   A multitude of studies have highlighted the potential of radiomics
   models as a promising tool for the diagnosis, management, and
   prognostication across various cancer types, including ovarian cancer,
   hepatocellular carcinoma, and meningioma^[117]42–[118]44. Several
   studies have specifically focused on constructing radiomics models to
   predict the survival of glioma patients^[119]45,[120]46. Li et
   al.^[121]47 established a prognostic radiomics model for patients with
   glioma, which demonstrated that patients with high radiomic feature
   scores had shorter OS. Similarly, our radiomics model, incorporating
   texture features, first order features, and shape features from
   radiomic images, consistently exhibited discriminatory performance to
   predict OS in HGG patients. According to Kaplan–Meier curves, the
   median OS of RS^low group was 21.6 months, which was 6.7 months longer
   than that of RS^high group. These findings collectively demonstrate the
   fantastic prognostic power of our radiomics models.

   While our radiomics models demonstrated strong performance in
   predicting CD44 expression and OS in patients with HGG, several
   limitations of the present research provide orientations for future
   study. Firstly, the research was an exploratory study, further
   validation study should be reported. In addition, the present study was
   based on MRI enhanced T1WI images, and future studies should consider
   incorporating additional functional MRI sequences. Moreover, the
   dataset used in this study was obtained from public database and the
   sample capacity was relatively small. Multicenter prospective studies
   are still needed to combat bias and enhance the generalizability of our
   findings.

Conclusion

   Elevated CD44 expression was found to be strongly associated with
   shorter OS in patients with HGG. Our radiomics model, referred to as
   RS, demonstrated the ability to accurately discern the level of CD44
   expression and predict OS in patients with HGG. Moreover, RS showed a
   significant correlation with some EMT related genes. This non-invasive
   radiomics model exhibited reliable and feasible performance, offering a
   potential powerful tool for predicting outcomes and assisting clinical
   decision-making in the era of individualized precision medicine and
   target therapy.

Electronic supplementary material

   Below is the link to the electronic supplementary material.
   [122]Supplementary Material 1^ (2.2MB, tif)
   [123]Supplementary Material 2^ (17.5KB, docx)
   [124]Supplementary Material 3^ (2.5MB, xlsx)
   [125]Supplementary Material 4^ (13KB, xlsx)
   [126]Supplementary Material 5^ (12.9KB, xlsx)
   [127]Supplementary Material 6^ (337.2KB, tif)
   [128]Supplementary Material 7^ (13.9KB, docx)

Acknowledgements