This article is mentioned in:

Introduction

Primary central nervous system lymphoma (PCNSL) and high-grade glioma (HGG) are malignant brain tumors, with annual incidence rates of ~0.45 and 4 per 100,000 population, respectively, in the United State (1). The diseases can mimic each other in clinical, radiological and even pathological examinations (2,3); however, they require completely different management. Approximately 90% of PCNSL cases are diffuse large B-cell lymphomas (DLBCL). High-dose methotrexate-based polychemotherapy plus rituximab is currently the treatment of choice for CNS DLBCL (4). Conversely, the primary treatment modality for HGG is surgery, followed by chemotherapy and radiotherapy (5). Furthermore, prior to surgery, corticosteroids can be administered to decrease the symptomatic tumor-associated edema of HGG. As the management is completely different for these entities, a precise diagnosis is crucial.

PCNSL and HGG can display overlapping clinical, radiological and partial pathological features that cause challenges in making the differential diagnosis between them. Firstly, both entities show a predilection for males and older patients (>65 years old) (1), commonly involving the deep white matter and the corpus callosum (6,7). The symptoms are associated with the involved areas of the brain, such as focal neurological deficits, neurocognitive impairment, seizures and signs of elevated intracranial pressure (headaches, nausea and vomiting) (6,8). Secondly, PCNSL and glioblastoma multiforme (GBM) can mimic each other in both imaging and histological examinations. For example, a case report (9) showed that CNS DLBCL in an immunocompromised patient was misdiagnosed as GBM by ring-like enhancement on contrast-enhanced T1-weighted imaging (CE-T1WI) and obvious edema in T2-fluid-attenuated inversion recovery (T2-FLAIR) images. This case was also misdiagnosed as GBM on intraoperative frozen section examination showing coagulation necrosis and atypical cell infiltration. DLBCL was recorded as the final diagnosis until postoperative formalin-fixed, paraffin-embedded (FFPE) tissue examination accompanied by immunohistochemistry (IHC) staining was performed. Similarly, another case report (10) mentioned that GBM was also nearly misdiagnosed as CNS DLBCL on magnetic resonance imaging (MRI) with and without gadolinium-based contrast, which showed little surrounding edema and mild diffusion restriction. Flow cytometry of the cerebrospinal fluid (CSF) also revealed an abnormal B-cell population, providing further evidence for the diagnosis of lymphoma. However, a stereotactic biopsy of the brain mass confirmed the diagnosis of GBM.

Pathological changes in FFPE tissues accompanied by IHC staining are acknowledged as the gold standard for the diagnosis of brain tumors. Microscopically, PCNSL typically exhibits high cellularity and a diffuse growth pattern. Large central areas of geographical necrosis are usually accompanied by variable to frequent perivascular lymphoma islands from central to peripheral lesions. Astrocytic and microglial proliferation, as well as inflammatory reactions, are often observed in the surrounding parenchyma. GBM also presents with large central necrotic areas surrounded by accumulated pleomorphic glial tumor cells in the periphery. Compared with PCNSL, HGG displays microvascular proliferation, which is usually most marked around necrotic foci and in the peripheral zone of infiltration. In some cases, there are isolated tumor cells or islands infiltrating the parenchyma rather than distinct masses, or the tumor cells show obvious heterogeneity, especially in stereotaxic needle biopsies; thus, further IHC staining is needed to make a differential diagnosis (11).

Stereotactic biopsy is sometimes inconclusive after the use of corticosteroids, which can induce rapid tumor shrinkage and false-negative results (12), especially with limited specimens. In one study, the rate of occurrence of false-negative biopsies after <1 week of steroid treatment was 33.3%, increasing to 57.1% after >1 week of steroid treatment (2). A reasonable interval of at least 2–4 weeks after withholding steroid usage is recognized as necessary prior to brain biopsy (2,13–15), yet this causes a delay in the diagnosis of PCNSL. Furthermore, stereotactic biopsy is not always feasible in patients (13). Thus, imaging, liquid analysis and molecular tests are quite meaningful for prompt diagnosis and can provide circumstantial evidence for pathological examinations, especially with limited samples. Although independent diagnostic research progress for PCNSL and HGG is abundant, the present narrative review only includes the progress from comparative diagnostic studies between PCNSL and HGG, with the purpose of building a liaison between radiology and pathology, as well as to provide more options for clinicians to improve the accuracy and speed of the decision-making process.

Radiological progress in PCNSL and HGG

Conventional MRI

According to the National Comprehensive Cancer Network guidelines (16), MRI with and without contrast serves as the first-line recommendation for the differentiation of PCNSL and HGG. CNS DLBCL typically presents as iso- to hypointense on unenhanced T1-weighted MRI (17), as iso- to hyperintense on T2-weighted imaging (T2WI) and as homogeneous enhancement on CE-T1WI (4). By contrast, HGG usually presents with central hypointensity similar to CSF and a marginal slightly high signal on T2WI, and as a heterogeneous, ring-enhancing lesion (5). Edema around the HGG lesion can be clearly observed on the T2-FLAIR sequence. As aforementioned, signal changes in these conventional MR sequences can often reflect some histological information, such as edema, hemorrhage and necrosis. However, there are considerable overlapping appearances between PCNSL and HGG in atypical cases, which cause confusion in conventional MRI, especially in cases of immunodeficiency and after treatment. For example, GBM with homogeneous enhancement but no apparent necrosis may mimic PCNSL, while atypical PCNSL with heterogeneous enhancement and central necrosis may mimic GBM (18). However, advanced MRI techniques can provide more evidence for the differential diagnosis.

Cellularity and diffusion-weighted imaging (DWI)

DWI calculates the apparent diffusion coefficient (ADC) values to quantify the diffusion of unbound extracellular water molecules (19). Differences in cellularity between PCNSL and GBM can be assessed using DWI. High cellular density and a large nuclear to cytoplasmic ratio lead to the reduction of water diffusion (19), which causes increased signal shadows and a decreased ADC value (20), differentiating untreated PCNSL from GBM (21–24). Different ADC parameters have been discussed for the purpose of differential diagnosis, including ADCmin, rADCmean, and ADC5% [the lowest and mean ADC value obtained from placed regions of interest was defined as the minimum ADC (ADCmin) and mean ADC (ADCmean), respectively. rADCmean=(ADCmean of the tumor)/(ADCmean of the normal appearing white matter of the contralateral hemisphere). Percentile values are the Nth percentile from the ADC histograms that reflect the distribution of ADC values in the region of interest. ADC5% is the fifth percentile from the ADC histograms that reflect the distribution of ADC values in the region of interest]. All ADC parameters are lower in PCNSL compared with those in GBM. A previous study (25) demonstrated a cutoff value of 0.68×10−3 mm2/sec for ADC5% to differentiate CNS DLBCL and GBM, with a sensitivity of 100%. In addition, molecular types of tumors may affect ADC values. For example, MYC proto-oncogene bHLH transcription factor (MYC) and BCL2 apoptosis regulator (BCL2) gene rearrangements in CNS DLBCL lead to lower rADCmin and rADCmean values, while isocitrate dehydrogenase [NADP(+)]1 (IDH1) mutation in GBM leads to higher values, when compared with wild-types (18). Methylation of the O(6)-methylguanine-DNA methyltransferase (MGMT) promoter in GBM may lead to higher ADC values with lower perfusion (26). Epidermal growth factor receptor variant III mutant GBM showed a higher cell density, leading to increased perfusion and a lower ADC value (27).

Combined conventional MRI and DWI can more accurately reflect histologically related information for the differential diagnosis. Fig. 1 presents representative combined images from the Department of Radiology of China-Japan Union Hospital of Jilin University (Changchun, China). For example, typical images of a representative DLBCL case occurring in the left temporal lobe are shown in Fig. 1. The lesions were classified as isointense on T1WI and T2WI, as high intensity on FLAIR, surrounded by finger-like edema limited diffusion on DWI, and as uniform with obvious enhancement on CE-T1W1. Representative images from a GBM case are also shown in Fig. 1. The lesions occurred in the left temporal lobe, with a low signal on T1WI and a high signal surrounded by finger-like edema on T2WI. The lesions were isointense on FLAIR, and showed restricted diffusion on DWI and garland-like enhancement on CE-T1W1. However, lymphoma often presents with atypical images, not only in conventional MRI, but also in DWI. Fig. 1 shows a representative case in which the lesions, which occurred in the posterior fossa and were finally diagnosed as DLBCL by pathological examination, presented as isointense on T1WI and T2WI, as slightly hyperintense on FLAIR, with no surrounding edema and no restricted diffusion on DWI, and as obvious uniform enhancement on CE-T1WI. Thus, more parameters were needed for the differential diagnosis.

Figure 1. Original magnetic resonance imaging of central nervous system lymphoma or high-grade glioma. Typical images of DLBCL are shown in the top row. The lesion displays as isointense on T1WI and T2WI, as high intensity on FLAIR surrounded by finger-like edema, with limited diffusion on DWI, and with uniform and obvious enhancement on CE-T1W1. Typical images of GBM lesions are shown in the middle row. The lesion displays a low signal on T1WI, a high signal surrounded by finger-like edema on T2WI, as isointense on FLAIR, with no restricted diffusion on DWI, and as garland-like enhancement on CE-T1W1. Atypical images of multiple DLBCL neoplastic lesions are shown in the bottom row. The lesions display as isointense on T1WI and T2WI, as slightly hyperintense on FLAIR with no surrounding edema, with no restricted diffusion on DWI and with obvious uniform enhancement on CE-T1WI. Red lines surround and indicate the lesion in each image. DLBCL, diffuse large B-cell lymphoma; CE-T1WI, contrast-enhanced T1-weighted imaging; FLAIR, fluid-attenuated inversion recovery; DWI, diffusion-weighted imaging; GBM, glioblastoma.

Angiogenesis and dynamic susceptibility contrast (DSC) MRI

DSC MRI is perfusion MRI using contrast medium injection. DSC MRI tracks the T2 weighted signal to the image and calculates perfusion metrics (28). Differences in angiogenesis between PCNSL and HGG can be reflected by cerebral blood flow (CBF) and cerebral blood volume (CBV). Compared with HGG, PCNSL displays lower CBF and CBV values due to the absence of neovascularization (29). Neska-Matuszewska et al (30) reported that max rCBV [max rCBV=(CBVmax of the tumor)/(CBVmax of the normal appearing white matter of the contralateral hemisphere)] demonstrated a high accuracy of 98.5% in differentiating 16 CNS B-cell lymphomas from 20 GBMs and 20 metastases. Arterial spin labeling (ASL) can be used to detect CBF values by non-invasively labeling blood without injection of contrast medium (31). It is suitable for patients who cannot undergo the injection of a contrast agent. A meta-analysis showed that DSC had a higher diagnostic accuracy than ASL in differentiating HGG from PCNSL (31). However, the accuracy of CBV decreases due to damage to the blood-brain barrier (BBB) from both HGG and PCNSL, which causes contrast agent leakage (32). A preload contrast dose is used to minimize the effects of leakage (19). After a preload injection, Chaganti et al (33) used a mean rCBV of 2.68 to differentiate 11 PCNSLs from 15 HGGs, with an area under the curve (AUC) of 1.000.

BBB damage and DSC or dynamic contrast-enhanced (DCE) MRI

Serious BBB damage can be analyzed by percentage signal recovery (PSR) and volume transfer constant (Ktrans), which can compensate for the deficiency of CBV. In the study by Cindil et al (32), both parameters were reported to be higher in PCNSL than in HGG. The PSR was calculated from the time-signal curve of DSC MRI. PSR represents a complex combination of tissue microstructure and hemodynamic effects, such as blood flow, vascular permeability and cellular geometry (34). PSR is a promising parameter that performs better than rCBV in both a PSR-optimized protocol without preload (AUC, 0.979) (32) and a CBV-optimized protocol with preload (AUC, 0.830) (35). Ktrans is a parameter of permeability calculated from T1-weighted signal curves and is obtained from DCE MRI (28). It is still controversial to evaluate the effect of the Ktrans parameter for the differentiation between PCNSL and HGG. One previous study (21) reported no significant difference in the distribution of Ktrans between CNS DLBCL and GBM. However, another study by Lu et al (22) revealed that a cutoff of 0.187 for Ktrans reached a high AUC of 0.852 for differentiating CNS DLBCL from GBM.

Microhemorrhage, calcification and neovascularity, and susceptibility-weighted imaging (SWI)

Differences in microhemorrhage, calcification and neovascularity (veins) between CNS DLBCL and GBM could be reflected by the intratumoral susceptibility signal (ITSS) on SWI (36). A higher ITSS presents more dot-like foci of susceptibility within the tumor. In one study, the ITSS was significantly higher in GBM than in B-cell PCNSL, and performed well in the differentiation of PCNSL from GBM, with an AUC of 0.800 (37). In addition, a recent study (18) reported lower ITSS and higher rSWI values [rSWI=(mean SWI of the tumor)/(mean SWI of the matching contralateral normal-appearing white matter)] in IDH1-mutant GBM compared with those in wild-type GBM, suggesting that it would be difficult to distinguish B-cell type PCNSL from IDH1-mutant GBM by ITSS or mean rSWI value.

Metabolism and MR spectroscopy (MRS) or positron emission tomography (PET)/computed tomography (CT)

The alterations in metabolites such as lipid and myo-inositol (mIns) within the tumor could be detected by MRS. Lipid levels are associated with necrosis and the activation of lymphocytes and macrophages, which are significantly higher in PCNSL than in GBM (8,38). The mIns level reflects the expression of inositol-3-phosphate synthase 1 (ISYNA1), which is the rate-limiting enzyme of the first step in the biosynthesis of mIns. Higher ISYNA1 expression in GBM leads to a higher mIns level (39).

Different uptake of radiotracers between PCNSL and HGG is presented by the tumor to normal contralateral cortex activity (T/N) ratio and standardized uptake value (SUV). The SUV and T/N ratio are semiquantitative indicators of radiotracer utilization. Most clinical centers use a radiolabeled glucose analog such as [18F] fluorodeoxyglucose (FDG) as the tracer (19). Other tracers such as 11C-methionine also show potential value in differentiating PCNSL from GBM; however, 11C-methionine has limited value in the clinical routine due to its short half-life (20 min) (40). The SUV of FDG is significantly higher in CNS lymphoma than in GBM due to the higher tumor cell density with elevated glycolytic metabolism (41). Nakajima et al (25) reported that a cutoff SUVmax value of 9.35 performed well in differentiating CNS DLBCL from GBM, with an AUC of 0.933. Similarly, Zhou et al (42) reported an AUC of 0.910 for differentiating 40 PCNSLs from 52 GBMs, with a higher SUVmax cutoff value of 13.77. However, some atypical PCNSLs show low FDG uptake, which is closely associated with the negative expression of mutated melanoma-associated antigen 1 protein (43).

Preferred radiology workflow for the differential diagnosis of PCNSL from HGG

Details of all radiological parameters and case numbers of PCNSL and GBM involved in a series of studies are summarized in Table SI. Marked differences between PCNSL and HGG were observed in DWI, DSC/DCE and 18F-FDG-PET/CT, which corresponded to the pathophysiological features of the two entities. As shown in Fig. 2, PCNSL presents with a diffuse growth pattern, obvious BBB damage and high cellularity associated with elevated glycolytic metabolism, which correspond to low ADC values in DWI, increased PSR values in DSC MRI, increased Ktrans values in DCE MRI and high FDG uptake on PET/CT. Furthermore, the feature of microvascular proliferation of HGG can be reflected by an increased CBV value in DSC.

Figure 2. Association between pathology and radiology in the differential diagnosis of PCNSL and HGG. PCNSL shows high cellularity, a diffuse growth pattern and BBB damage, and a high glycolytic metabolism, which correspond to low ADC values in DWI, increased PSR values in DSC MRI, increased Ktrans values in DCE MRI and high FDG uptake on PET/CT. By contrast, obvious angiogenesis of HGG can be reflected by an increased CBV value in DSC MRI. HGG, high-grade glioma; PCNSL, primary central nervous system lymphoma; BBB, blood-brain barrier; ADC, apparent diffusion coefficient; DWI, diffusion-weighted imaging; PSR, percentage signal recovery; Ktrans, volume transfer constant; MRI, magnetic resonance imaging; DSC, dynamic susceptibility contrast; DCE, dynamic contrast-enhanced; FDG, [18F] fluorodeoxyglucose; PET/CT, positron emission tomography/computed tomography; CBV, cerebral blood volume.

Progress of radiomics and machine learning

Radiomics analyses provide promising evidence for a differential diagnosis, especially in tumors with high heterogeneity (44). Radiomics includes feature-based and deep learning-based components. Feature-based radiomics uses a series of mathematically predefined features that are typically extracted from a segmented region (45). Deep learning-based radiomics utilizes artificial neural networks that imitate the vision of humans and automatically extract high-dimensional features from the input images at different levels of scaling and abstraction (45).

Feature-based radiomics

The diagnostic value of image features extracted from different MRI types was researched for feature-based radiomics. Priya et al (46) found that features extracted from CE-T1WI could distinguish between 46 PCNSLs and 97 GBMs, with an AUC of 0.924. Different combinations of conventional MRI such as T1WI, T2WI, CE-T1WI and FLAIR also performed well using feature-based radiomics (47–49). MacIver et al (50) reported the high performance of an ADC map for diagnosing 48 PCNSLs and 42 GBMs (AUC, 0.880). Mehrnahad et al (51) also concluded the usefulness of feature-based radiomics derived from an ADC map in differentiating 57 GBMs from 25 PCNSLs. The diffusion condition of tumors in the ADC map was combined with CE-T1WI and analyzed by feature-based radiomics. The combination was as effective as radiologists, with an AUC of 0.946 when differentiating 14 PCNSLs from 28 GBMs (52). In one study, a radiomics model combining CE-T1WI and ADC maps (AUC, 0.935) performed better than veteran radiologists (AUC, 0.923-0.945) for distinguishing between 26 PCNSLs and 22 GBMs (53). Kim et al (54) used multi-parametric MRI (T2WI, CE-T1WI and ADC) for differentiating PCNSL and GBM, with an AUC of 0.956. Combining CE-T1WI, ADC and FLAIR also showed high accuracy (AUC, 0.977) (55). Conventional MRI could also be combined with perfusion-weighted MRI. In the study by Nakagawa et al (56), compared with the results of two radiologists, a model that extracted features from T2WI, CE-T1WI, an ADC map and an rCBV map provided a better performance for diagnosing 25 PCNSLs and 45 GBMs.

Deep learning-based radiomics

Deep learning-based radiomics also perform well in the differential diagnosis. McAvoy et al (57) identified the high accuracy of convolutional neural networks (CNNs) based on CE-T1WI for the differential diagnosis between 24 CNS-DLBCLs and 35 GBMs. In a larger population (92 PCNSLs and 97 GBMs), deep learning radiomics with data enhancement performed better than two neuroradiologists (58). Another group also reported CE-T1W1-based multiparametric CNNs and showed that it had a similar accuracy to radiologists (accuracy, 0.899; P=0.886) for differentiating between 136 PCNSLs and 153 GBMs (59). For differentiating 14 atypical GBMs from 11 PCNSLs, the CE-T1WI-based deep learning model also performed well (60).

A systematic review (44) announced the suboptimal quality of radiomics studies in lymphoma and suggested further research before extensive clinical use. Most radiomics studies are retrospective without large cohorts. Additionally, there are no standard methods for image acquisition, model generation and evaluation. For example, a cohort study with small sample sizes displayed a lower AUC of CNNs (0.486) based on CE-T1WI and ADC maps compared with that of feature-based radiomics (0.947) or radiologists (0.913-0.932) to distinguish 14 PCNSLs from 28 GBMs (61). However, compared with feature-based radiomics, deep learning-based radiomics requires less time in manual feature extraction or tumor segmentation. Another radiomics model, in which the perfusion information was captured by deep learning-based radiomics, showed a similar performance to rCBV in differentiating 15 PCNSLs from 28 GBMs (62). Metabolism assessed by FDG-PET-based radiomics also showed potential value in differentiating between 24 CNS DLBCLs and 53 GBMs, with an AUC of 0.971-0.998 (41).

In summary, radiomics remains a potential tool for the discrimination between PCNSL and HGG. Details of all radiomics models and case numbers of PCNSL and GBM involved in a series of studies are summarized in Table SII.

Progress of CSF and blood analysis in PCNSL and HGG

Cytology and flow cytometry

CSF sampling is an additional choice other than biopsy if it is safe and does not delay the diagnostic process or treatment (7). CSF analysis includes flow cytometry and cytology, and may consider gene rearrangements (16). In cytology, GBM cells have hyperchromatic nuclei and a variable amount of cytoplasm, appearing singly or arranged in small cohesive groups, while PCNSL usually consists of a monotonous population of dyshesive large lymphoid cells (63). One study mentioned that flow cytometry of CSF improves the sensitivity up to 2–3 times more than cytology analysis (64). Furthermore, another research group (65) defined the sensitivity of cytology and flow cytometry as 13.3 and 23.3%, respectively, by analyzing the CSF of 30 patients with PCNSL. However, studies confirmed considerable diagnostic delays of PCNSL that may take up to several weeks due to lengthy cytology or flow cytometry analysis of CSF, especially under conditions of limited or rare tumor cells in CSF specimens where more procedures will be required, and detection time will be further increased (13,66). The detection of biomarkers in CSF and blood may improve the diagnosis. Diagnostic biomarkers include proteins, RNA, DNA and extracellular vesicles (EVs).

Protein

Proteins in CSF can be analyzed by latex agglutination turbidimetric immunoassay (LATIA), enzyme-linked immunosorbent assay (ELISA) and targeted proteomics assays such as selected reaction monitoring (SRM). Some proteins, such as C-X-C motif chemokine ligand 13 (CXCL13), interleukin-10 (IL-10), β2-microglobulin (B2M), soluble IL-2 receptor (sIL-2R), apolipoprotein C-II (APOC2), glycoprotein non-metastatic melanoma protein B (GPNMB), and V-set and immunoglobulin domain-containing protein 4 (VSIG4), are considered promising protein biomarkers for PCNSL in the CSF (67). CXCL13 and IL-10 are mediators of the migration and growth of B cells, respectively. B2M is a component of the major histocompatibility complex class I molecules. sIL-2R is a receptor that is mostly released by regulatory T cells (68). APOC2 is an exchangeable apolipoprotein that activates lipoprotein lipase (69). GPNMB is an endogenous type 1 transmembrane glycoprotein (70). VSIG4 is a phagocytic receptor that negatively regulates T-cell proliferation and IL-2 production (71). All these proteins may be upregulated in the CSF of patients with PCNSL compared with patients with GBM (71–75). The combination of different biomarkers can improve diagnostic performance. Maeyama et al (72) proposed a multi-marker prediction algorithm that was effective (AUC, 0.994) in distinguishing 32 PCNSLs from 21 GBMs and 51 other brain lesions by incorporating the detection of CXCL13, IL-10, B2M and sIL-2R in the CSF. In the study, only B2M was measured by LATIA, while the remaining biomarkers were analyzed by ELISA. Combined analysis of APOC2, GPNMB and VSIG4 proteins in CSF was effective (AUC, 0.953) in distinguishing 28 PCNSLs from 7 GBMs 3 astrocytomas using the SRM method (71).

DNA

Alterations in tumor-specific genes and methylation of circulating tumor DNA (ctDNA) can be detected by gene-targeted PCR for the purpose of differential diagnosis, with high sensitivity (76). Cell-free DNA (cfDNA) is DNA shed from the cell into the body fluids. A portion of cfDNA is derived from tumor cells, which is referred to as ctDNA (71). The CSF sampling serves as a better choice for the detection of mutations compared with plasma, which contains less cfDNA or ctDNA than CSF due to the BBB (76–78). Myeloid differentiation primary response gene 88 (MYD88) encodes a cytosolic adapter protein that plays a central role in innate and adaptive immunity. CD79B encodes the B-cell antigen receptor complex-associated protein β chain. The alterations of L265 in MYD88 and Y196 in CD79B are detected in recurrent lymphoid tumors. A large cohort study (79) detected mutations in MYD88 L265 and CD79B Y196 in 71.7% and 64.2% of CNS DLBCL cases, respectively, but no MYD88 or CD79B mutations were detected in GBM. Hiemcke-Jiwa et al (80) used the droplet digital PCR method to detect MYD88 mutation in cfDNA and found that 8 out of 11 CSF specimens from PCNSL were positive, including 2 unknown PCNSL samples and 6 CNS DLBCL samples, while all CSF samples from 3 GBMs were negative for the mutation. Combining gene mutations with protein markers can improve diagnostic performance. Ferreri et al (81) combined MYD88 mutational status (assessed by TaqMan-based PCR) with IL-10 levels in CSF and differentiated 36 PCNSLs from 106 other CNS diseases (10 GBMs), with 94% sensitivity and 98% specificity. Furthermore, certain methylated DNA markers can provide valuable diagnostic clues for differentiation. Wang et al (77) detected the methylation of MGMT promoter in ctDNA in 18 out of 28 CSF samples from HGG using methylation-specific PCR. Recently, Downs et al (82) identified methylation markers cg054 and SCG3 in plasma via a novel method of tailed amplicon multiplexed-methylation-specific PCR, with a high level of accuracy to distinguish PCNSLs from other CNS tumors.

RNA

Different physiological and pathological processes between PCNSL and GBM can also be reflected by extracellular RNAs (exRNAs), such as small non-coding RNA (ncRNA) (83), which freely exist in body fluids or concentrated in carriers such as EVs (84). MicroRNAs (miRNAs/miRs) are snRNAs of 18 to 24 nucleotides in length (85). EVs are small lipid bilayer-enclosed vesicles containing proteins, lipids and nucleic acids that release from cells (86). A number of exRNAs, such as free miRNA (miR-15b and miR-21) and RNA in EVs (RNU6-1) have been used to differentiate between PCNSL and GBM in previous studies (87). miR-15b is important in glioma carcinogenesis for regulating cell cycle progression (88). One study reported higher miR-15b levels in the CSF of 10 gliomas compared with those in 23 PCNSLs and 17 other CNS diseases, with an AUC of 0.960 (89). miR-21 is one of the most highly expressed miRNAs and mainly targets the phosphatase and tensin homologue (PTEN) gene (90). A previous study showed that PCNSL cases displayed relatively higher miR-21 levels compared with those in patients with GBM and healthy individuals. miR-21 levels are higher in the CSF (89) and serum (87) of B-cell type PCNSL than those of GBM. RNU6-1 is an snRNA that is negatively regulated by the PTEN pathway (91). More alterations of the PTEN pathway in GBM than in PCNSL leads to higher RNU6-1 expression (92). A study showed that high levels of RNU6-1 in EVs derived from serum helped differentiate between 18 GBMs and 12 PCNSLs, with an AUC of 0.700 (92). Overall, the limited studies on miRNAs with regard to distinguishing PCNSL from HGG have presented heterogeneous results without consistent expression variation across cohorts due to the lack of standards for liquid collection, RNA extraction, RNA sequencing and statistical analysis. A large cohort with comparative research of the aforementioned miRNAs will be meaningful for potential clinical application.

Preferred liquid biomarkers for the differential diagnosis of PCNSL from HGG

Details of all of the liquid parameters and case numbers of PCNSL and GBM involved in a series of studies are summarized in Table SIII. Among these data, as shown in Fig. 3, examinations of CXCL13 and IL-10 showed more supportive evidence. Since DLBCL accounts for most PCNSL cases, MYD88 L265 and CD79B Y196 in the CSF are meaningful for the diagnosis of PCNSL. Other promising evidence, such as an increased miR-21 level in the CSF and plasma, increased cg0504 and SCG3 levels in the plasma, and increased sIL-2R, β2-MG, APOC2, GPNMB and VSIG4 levels in the CSF, are also suggestive of PCNSL, yet further evidence is required.

Figure 3. Prospects in liquid biopsy biomarkers of the CSF and plasma. Among these markers, CXCL13 and IL-10 in the CSF show more supportive evidence. Since diffuse large B-cell lymphoma accounts for most PCNSL cases, MYD88 L265 and CD79B Y196 in the CSF are meaningful for the diagnosis of PCNSL. Other promising evidence, such as increased miR-21 level in the CSF and plasma, increased cg0504 and SCG3 levels in the plasma, and increased sIL-2R, β2-MG, APOC2, GPNMB and VSIG4 levels in the CSF, also suggest PCNSL. CSF, cerebrospinal fluid; CXCL13, C-X-C motif chemokine ligand 13; IL-10, interleukin-10; B2M, β2-microglobulin; sIL-2R, soluble IL-2 receptor; APOC2, apolipoprotein C-II; GPNMB, glycoprotein non-metastatic melanoma protein B; VSIG4, V-set and immunoglobulin domain-containing protein 4; PCNSL, primary central nervous system lymphoma; miR, microRNA; LATIA, latex agglutination turbidimetric immunoassay; ELISA, enzyme-linked immunosorbent assay; SRM, selected reaction monitoring; FCM, flow cytometry; cytopath, cytopathology; RT-qPCR, reverse-transcription-quantitative PCR; ddPCR, droplet digital PCR; TAM-MSP, tailed amplicon multiplexed-methylation-specific PCR; MYD88, myeloid differentiation primary response gene 88.

Promising combination studies of multiple analyses

Combining different detection methods or parameters may provide more accurate and robust support for diagnosis (Table SIV). Combined technologies, including MRI and DWI, have most commonly been used (22,23,93). A previous study (22) combined rADC with Ktrans to distinguish 18 CNS DLBCLs from 42 GBMs, and this method performed better than using each parameter alone. Makino et al (23) designed a two-step decision tree with an rCBVmax of 4 and ADCmin of 1. The combination facilitated differentiation between 33 PCNSLs and 54 GBMs. Another study (93) showed that the combination of mean and maximum CBV, mean and maximum selective ADC, and rCBVmean could obtain 100% accuracy for discrimination between 37 PCNSLs and 37 GBMs. A corporation analysis of diffusion and susceptibility also showed a good performance (18). Ozturk et al (18) combined rSWI with rADCmean to differentiate 31 B-cell type PCNSLs with BCL2 and MYC rearrangements from 57 atypical GBMs without visible necrosis. The combination improved the diagnostic performance to an AUC of 0.936. Saini et al (37) used ADC, corrected rCBV, back flux exchange rate and ITSS scores to perform a multiparametric assessment of 30 B-cell PCNSLs and 70 GBMs. The study reported the good performance of this model, with an AUC of 0.920. A combination of ADC values and biomarker analysis has also been used in research (94), including the average ADC, and CXCL13 and IL-10 levels in the CSF, with a better performance than any single variable model. Recently, a combination of 18F-FDG-PET and ASL was also introduced to differentiate PCNSL from GBM, which achieved a better performance than either technique alone (95). Yet, to the best of our knowledge, there is no standard multiparametric model that has been assessed prospectively in a large population. Further studies are needed.

Conclusions and future prospects

Accurate diagnosis is a prerequisite for the precise treatment of PCNSL and GBM. Multidisciplinary participation, including clinicians, radiologists and pathologists, is meaningful for improving the accuracy and speed of the decision-making process. However, most of the radiological and pathological markers discussed within the present review are from retrospective small-scale and heterogeneous cohorts. Thus, current evidence from these studies shows the limited quality of radiological and liquid biopsy, which is unable to replace the need for histology examination.

Although stereotactic biopsy is an optimal choice for diagnosis, it is not always conclusive. Otherwise, surgical resection may be another option, not only for providing enough pathological specimens but also for significantly improving the overall survival and progression-free survival of certain patients with PCNSL (96,97). A recent systematic review highlighted the possible benefit of cytoreductive surgery in PCNSL (98). Another recent study reported that resection significantly prolonged the survival of patients <70 years old with superficial solitary lymphoma lesions (99). However, when lymphoma lesions are deeply seated, it is frequently not amenable to resect them, and biopsy will be a safer option (99,100).

Although radiological examination, and CSF and blood analysis are useful for the diagnosis of PCNSL, it is not always practical in clinical due to the economic and staff/equipment time aspects. Thus, as shown in Fig. 4, based on the literature, we propose an optional procedure using multiparametric imaging technologies and liquid biopsies to improve the diagnostic performance and management of CNS neoplasms, especially PCNSL. More evidence should be collected in the future for the clinical management of PCNSL and in order to make the current diagnostic workflow more reasonable and practical.

Figure 4. Integrated radiological-pathological evidence for the differential diagnosis of PCNSL from HGG. For patients with suspected CNS lesions, conventional MRI is preferred before stereotactic biopsy. If a biopsy is not tolerated or the specimen is limited, a selective combination of advanced MRI (including DWI, ASL/DSC MRI and DCE MRI), PET-CT and liquid biopsy will provide circumstantial evidence for the differential diagnosis. CNS, central nervous system; MRI, magnetic resonance imaging; DWI, diffusion-weighted imaging; ASL, arterial spin labeling; DSC, dynamic susceptibility contrast; DCE, dynamic contrast-enhanced; MRI, magnetic resonance imaging; MRS, magnetic resonance spectroscopy; PET/CT, positron emission tomography/computed tomography; CSF, cerebrospinal fluid; PCNSL, primary central nervous system lymphoma; HGG, high-grade glioma.

Supplementary Material

Supporting Data

Supporting Data

Supporting Data

Supporting Data

Acknowledgements

Not applicable.

Funding

This study was supported by the Young Scientists Fund of the National Natural Science Foundation of China (grant no. 81700198), the Science and Technology Development Project of Jilin Province (grant no. 20190701064GH), the Open Project of Key Laboratory of Tumor Immunology and Pathology (Army Medical University), Ministry of Education (grant no. 2018jsz101) and the College Student Innovation and Entrepreneurship Training Program of Jilin University (grant no. 202010183X453).

Availability of data and materials

Not applicable.

Authors' contributions

XJW developed the concept, designed the study and revised the manuscript. LMC collected the literature and wrote the draft. MCZ and YZ respectively reorganized the context in radiology and liquid biopsy. LMC, MCZ and YZ contributed equally. BJ and XMW respectively contributed suggestions on PET-CT and pathology. MCZ, YZ, BJ and LMC participated in writing or revising the review. All authors read and approved the final manuscript. Data authentication is not applicable.

Ethics approval and consent to participate

Not applicable.

Patient consent for publication

Written informed consent was obtained from the individuals for the publication of any potentially identifiable images or data included in this article.

Competing interests

The authors declare that they have no competing interests.

References