Annals of Nuclear Medicine

, Volume 25, Issue 6, pp 432–438

Does adding FDG-PET to MRI improve the differentiation between primary cerebral lymphoma and glioblastoma? Observer performance study

  • Keishi Makino
  • Toshinori Hirai
  • Hideo Nakamura
  • Ryuji Murakami
  • Mika Kitajima
  • Yoshinori Shigematsu
  • Rumi Nakashima
  • Shinya Shiraishi
  • Hiroyuki Uetani
  • Koya Iwashita
  • Masuma Akter
  • Yasuyuki Yamashita
  • Jun-ichi Kuratsu
Original Article

DOI: 10.1007/s12149-011-0483-1

Cite this article as:
Makino, K., Hirai, T., Nakamura, H. et al. Ann Nucl Med (2011) 25: 432. doi:10.1007/s12149-011-0483-1

Abstract

Objective

It is sometimes difficult to distinguish between primary central nervous system lymphomas (PCNSL) and glioblastoma multiforme (GBM). The aim of this study was to investigate whether the addition of 18F-2-fluoro-2-deoxy-d-glucose positron emission tomography ([18F]FDG-PET) and apparent diffusion coefficients (ADC) to conventional MRI improves diagnostic accuracy for distinguishing between PCNSL and GBM with similar MRI findings.

Methods

We used conventional- and diffusion-weighted MRI and FDG-PET scans of 21 patients with histologically confirmed brain tumors exhibiting similar MRI findings (PCNSL, n = 14, GBM, n = 7) in our observer performance study that consisted of 3 interpretation sessions. ADC and maximum standard uptake values (SUVmax) of the tumors were calculated. Three radiologists first interpreted conventional MRI (1st session), then they read images to which the ADC value had been added (2nd session), and finally they interpreted images supplemented with SUVmax (3rd session). Observer performance was evaluated using κ statistic and receiver operating characteristics analyses.

Results

The addition of ADC values to conventional MRI failed to improve the differentiation between PCNSL and GBM. The addition of SUVmax at the third session improved the diagnostic accuracy of all 3 readers and resulted in higher interobserver agreement; mean accuracy was 95% (range 93–100%). In one observer the accuracy of tumor differentiation was significantly improved at the third compared to the second session (p = 0.017).

Conclusions

In a selected group of PCNSL and GBM with similar MRI findings, the addition of quantitative FDG-PET to MRI may improve their differentiation. ADC measurement did not allow further discrimination.

Keywords

Brain tumors PET MRI 

Introduction

Primary central nervous system lymphomas (PCNSL) are extranodal malignant tumors that arise in the brain, spinal cord, leptomeninges, or eyes; they account for less than 5% of all primary CNS tumors. As the incidence of PCNSL has increased over the last two decades all over the world [1, 2, 3], there is a need to differentiate between PCNSL and glioblastoma multiforme (GBM), the most common primary brain tumor in adults. For the selection of appropriate treatment strategies, the accurate preoperative differentiation between PCNSL and GBM is imperative. For example, if the tumor is highly suspected of being PCNSL, stereotactic biopsy is usually recommended to confirm the diagnosis. If the tumor is highly suspected of being GBM, craniotomy would be chosen. Chemotherapy regimens are different for the two tumors.

It is often possible to differentiate between PCNSL and GBM on conventional MR images (MRI) alone. However, it can be difficult to distinguish between these tumors, especially when the PCNSL lesion is solitary and heterogeneously enhanced [4, 5, 6] or when the GBM lesion is homogeneously enhanced [7, 8]. Under these conditions additional imaging methods may be needed to improve diagnostic accuracy. Apparent diffusion coefficient (ADC) measurements in the tumor may be useful for the differentiation between PCNSL and high-grade glioma [9, 10, 11, 12]. 18F-2-fluoro-2-deoxy-d-glucose positron emission tomography (FDG-PET) has also been reported as a useful tool for differentiating PCNSL from high-grade glioma [13].

To our knowledge, there are no published data on the effect of adding ADC and FDG-PET to conventional MRI for the differentiation between PCNSL and GBM exhibiting similar MRI findings. Therefore, we studied whether the addition of ADC and FDG-PET to conventional MRI studies improves the diagnostic performance of radiologists charged with differentiating between PCNSL and GBM with similar MRI findings.

Materials and methods

Patients

Inclusion criteria for the observer performance study were (a) a PCNSL or GBM diagnosis histologically confirmed by biopsy or open surgery, (b) a solitary tumor with heterogeneous enhancement on conventional MRI for PCNSL, (c) a tumor with relatively homogeneous enhancement on conventional MRI for GBM, (d) no apparent body lesions on FDG-PET, (e) the availability for review of whole series of preoperative FDG-PET and conventional MR imaging data including ADC maps, (f) no corticosteroid and/or radiation therapy prior to PET and MRI study, (g) absence of diabetes mellitus, immunosuppression, or human immunodeficiency virus infection, and (h) an interval of 10 days or less between FDG-PET- and MRI studies. The presence of 2 or more enhanced lesions within a region of T2 prolongation was considered to represent a solitary lesion. Since GBM and PCNSL with typical MRI findings can be easily differentiated on conventional MRI scans, we chose tumors with following MRI findings: GBM with relatively homogeneous enhancement and PCNSL with a solitary mass and heterogeneous enhancement.

One experienced neuroradiologist (T.H.) and one experienced neurosurgeon (K.M.) evaluated the cases together. Between January 2003 and December 2008, these criteria were satisfied by 21 patients (13 males, 8 females, age range 37–81 years, mean 67 years); 14 presented with PCNSL and the other 7 with GBM. The total number of cases with GBM and PCNSL during the period were 123 and 60, respectively. The rate of the selected case with GBM was 7/123 (5.7%); that with PCNSL was 14/60 (23.3%).

The main location of PCNSL was the parietal lobe in 3 patients, the basal ganglia in 3, the thalamus in 3, the corpus callosum in 2, and the frontal and parietal lobes and cerebellum in one each; that of GBM was the temporal lobe in 3, and the basal ganglia, thalamus, corpus callosum, and frontal lobe in one each.

MRI

MRI scans were obtained on a 1.5-T superconducting system (Magnetom Vision; Siemens, Erlangen, Germany) until December 2006, thereafter we used a 3.0-T instrument (Magnetom Trio; Siemens). Conventional MRI and diffusion-weighted images (DWI) were acquired during the same procedure. For conventional MRI study we used a sagittal T1-weighted localizing sequence and axial T1-weighted, fast spin-echo T2-weighted, fluid-attenuated inversion-recovery (FLAIR), and triplanar contrast-enhanced T1-weighted sequences. For contrast-enhanced studies we intravenously (i.v.) injected gadopentetate dimeglumine (0.1 mmol/kg body weight; Magnevist; Bayer-Schering, Berlin, Germany). The parameters for T1- and T2-weighted and FLAIR sequences were: section thickness 5 mm, section gap 1 mm, matrix size 256 × 256–512, field of view (FOV) 220 × 220 mm.

DWI at the 2 MRI units was performed in the axial plane by using a single-shot echo planar imaging sequence; the parameters were: TR ms/TE ms 12000/100 for 1.5T and 3600/81 for 3T, diffusion gradient encoding in 3 orthogonal directions b = 0 and 1000 s/mm2, FOV 230 × 230 mm, matrix size 128 × 128 pixels, section thickness 5 mm, section gap 1 mm. DWI was performed before contrast-enhanced T1-weighted imaging.

The ADC values were calculated as: ADC = −[ln(S/S0)]/b, where S is the signal intensity (SI) of the region of interest (ROI) on 3 orthogonally oriented DWIs or diffusion trace images, S0 is the SI of the ROI on reference T2-weighted images, and b is the gradient b factor with a value of 1000 s/mm2. ADC maps were calculated on a pixel-by-pixel basis.

FDG-PET

For FDG-PET studies we used a PET/CT scanner (Discovery ST; GE Healthcare, Milwaukee, WI) for all 21 patients. All PET studies were performed as head- and full head-to-thigh scans to evaluate both the brain tumor and other body lesions. After a fast, lasting a minimum of 5 h, head FDG-PET scans were obtained 50 min after the i.v. injection of 3.7 MBq/kg of FDG; this was followed by head-to-thigh scanning. The reconstructed images had a slice thickness of 2 mm and a matrix of 128 × 128 for the head and of 4 mm and 144 × 144 for the body. Before PET we acquired unenhanced CT scans from the head to the upper thigh; the settings were: transverse 3.75-mm section thickness, 120 kVp, 108 mA, and 16.8-mm table speed. Images were reconstructed by ordered-subset expectation maximization two-dimensional interactive reconstruction; CT images were used to produce attenuation correction values for PET emission reconstruction and fused PET/CT presentation. Reconstruction images were converted to standard uptake value (SUV) images using the equation: SUV = activity at a pixel (kBq/cm3)/injection dose (MBq)/weight (kg).

Image analysis

PET/CT images were analyzed using PET/CT software (Xeleris, GE Healthcare). Several circular ROIs were placed over the tumor using information obtained by an experienced radiologist (R.N.) from contrast-enhanced MRI studies. Slices displaying maximum tumor activity were selected. Necrotic or cystic tumor portions on contrast-enhanced MRI scans were excluded from the ROI. The ROI area over tumors was at least 1.5 cm2. We used maximum SUV (SUVmax) among the SUVs obtained [13].

To assess the brain tumors based on imaging-, surgical-, and histopathologic data, one neuroradiologist (T.H.) and one neurosurgeon (K.M.) with 18 and 13 years of experience in brain MRI, respectively, consensually evaluated the entire series of MRI scans on a PACS workstation. They were cognizant of clinical and histopathologic data. They also consensually identified solid tumor components with contrast enhancement on conventional MRI scans and ADC maps. They carefully inspected the conventional MR images; ROIs were manually drawn on ADC maps in areas corresponding to the enhancing portion of the lesions. According to the previous reports [9, 11, 12], ROIs with diameter ranges from 0.7 to 1.2 cm were placed centrally within the largest solid-enhancing area of the tumor to avoid volume averaging with cysts and necrotic areas that might have influenced ADC values.

Observer performance study

Three radiologists (Y.S., R.M., M.K.) participated in the observer performance study. Each reader had 15–18 years of experience (mean 17 years) interpreting brain MRI. They were blinded to all clinical data including pathologic findings, but they were informed that the tumor was either PCNSL or GBM. They used a 5-point confidence scale; no limit was imposed on the reading time.

Each case was subjected to 3 interpretation sessions using a PACS workstation and a 21-in. monitor. At the first session the readers were provided with only the conventional MRI scan to evaluate morphologic features, SI, and lesional contrast enhancement patterns [4, 5, 6, 7]. Axial views of pre and postcontrast T1-weighted images, and T2-weighted and FLAIR images were evaluated. Each reader assessed the lesions using a 5-point confidence scale where 1 = definitely PCNSL, 2 = probably PCNSL, 3 = equivocal, 4 = probably GBM, and 5 = definitely GBM.

After the first interpretation, each reader performed a second interpretation where ADC maps and ADC values in the lesion were shown. Each reader was provided with the recommended optimal cutoff point for an ADC value and its accuracy, sensitivity, and specificity and re-assessed the lesion using the same 5-point confidence scale.

For the third interpretation, readers were shown the PET images and SUVmax measured in the lesion. The recommended optimal cutoff point for SUVmax and accuracy, sensitivity, and specificity data were provided and each reader re-assessed the lesion using the same 5-point confidence scale. After completing the third interpretation of individual cases, the readers proceeded to the first interpretation of the next case.

Statistical analyses

For each tumor type the mean and standard deviation (SD) of the ADC and SUVmax were calculated. These values of the solid-enhancing areas of the PCNSL and GBM were compared using the unpaired t test. ROC curves were generated from the ADC values and SUVmax of the 21 cases included in this study. Accuracy (Az), represented by the area under the ROC curve, and the cutoff point, sensitivity and specificity of the ADC and SUVmax were calculated and recorded. These data were used in the observer performance study. The difference in accuracy between the ADC alone and the SUVmax alone was evaluated for statistical significance by pairwise comparison of the ROC curves [14].

In each observer performance study, ROC analysis was performed using the 5-point confidence scale. Accuracy, represented by the area under the ROC curve, sensitivity, and specificity of each reader’s interpretation, with corresponding standard errors and variance, were calculated. Changes in each reader’s accuracy between the first and second interpretation session were evaluated for statistical significance by using pairwise comparison of the ROC curves [14].

Interobserver differences were assessed to establish the reliability of the interpretations. The degree of the interobserver difference between combinations of 2 readers was calculated with the chance-corrected κ statistic. κ statistic results were categorized as follows: a κ value greater than or equal to 0 but less than or equal to 0.20 = slight agreement, a κ value greater than or equal to 0.21 but less than or equal to 0.40 = fair agreement, a κ value greater than or equal to 0.41 but less than or equal to 0.60 = moderate agreement, a κ value greater than or equal to 0.61 but less than or equal to 0.80 = substantial agreement, a κ value greater than or equal to 0.81 but less than or equal to 1.00 = excellent agreement [15].

All statistical analyses were performed with standard statistical methods using a computer software package (MedCalc; MediSoftware, Mariakerke, Belgium). Differences of p < 0.05 were considered significant.

Results

ROC analysis of ADC and SUVmax

The distribution of the ADC values and SUVmax for PCNSL and GBM is shown in Fig. 1. The mean ± SD of the ADC values was 0.81 ± 0.21 × 10−3mm2/s (range 0.48–1.21 × 10−3mm2/s) for PCNSL and 0.94 ± 0.24 × 10−3mm2/s (range 0.69–1.39 × 10−3mm2/s) for GBM; the difference was not statistically significant. For SUVmax they were 16.76 ± 7.19 (range 7.9–30.5) for PCNSL and 8.24 ± 3.05 (range 4.0–12) for GBM. The mean SUVmax of PCNSL was significantly higher than of GBM (p < 0.01).
Fig. 1

Scatterplots of the ADC value (a) and SUVmax (b) measured in solid-enhancing PCNSL and GBM tumors. There was no significant difference between the mean ADC value of PCNSL and GBM. The mean SUVmax of PCNSL was significantly higher than of GBM (p < 0.01)

The ROC curve generated from ADC values and SUVmax in the 21 lesions is shown in Fig. 2. The optimal cutoff point for the ADC and SUVmax was 0.8 × 10−3mm2/s and 12, respectively. An ADC of 0.8 × 10−3mm2/s or greater suggested GBM, whereas an ADC of less than 0.8 × 10−3mm2/s suggested PCNSL. The accuracy (Az) of the ADC alone for lesion differentiation with a cutoff point of 0.8 × 10−3mm2/s was 0.68; sensitivity was 71.4 and specificity 64.3%. A SUVmax of 12 or greater suggested PCNSL; a value less than 12 suggested GBM. The accuracy (Az) of SUVmax alone for lesion differentiation with a cutoff point of 12 was 0.86; sensitivity was 100%, and specificity was 71.4%. There was no significant difference in the accuracy (Az) between SUVmax and the ADC value (p = 0.19).
Fig. 2

ROC curve generated from the ADC value and SUVmax measured in 21 PCNSL and GBM lesions. The solid line represents the ROC curve generated from the lesional ADC values; the dashed line the ROC curve generated from the SUVmax of the lesions. With an ADC value cutoff point of 0.8 × 10−3mm2/s, accuracy (Az) was 0.68, sensitivity was 71.4%, and specificity was 64.3%. With a SUVmax cutoff point of 12, Az was 86%, sensitivity was 100%, and specificity was 71.4%. There was no significant difference in accuracy (Az) between SUVmax and the ADC value (p = 0.19)

Observer performance study and interobserver agreement

As shown in Table 1, adding the ADC at the first interpretation session resulted in 2 readers achieving slightly higher accuracy in the discrimination between PCNSL and GBM; it was lower for the other reader. The mean accuracy for lesion differentiation was 74% (range 68–82%) at the first and 75% (range 61–89%) at the second session; the difference between the 2 sessions was not significant. Figure 3 shows the ROC curves generated from the probability of a PCNSL assignment by a representative reader.
Table 1

Accuracy of assigned probabilities of differentiating PCNSL from glioblastoma

Reader and interpretation

Accuracy of assigned probability

Value (%)

95% CI

p value

Reader 1

 Initial interpretation

82

0.59–0.95

 

 Second interpretation

89

0.68–0.98

0.50

 Third interpretation

100

0.84–1.00

0.13

Reader 2

 Initial interpretation

71

0.48–0.89

 

 Second interpretation

75

0.52–0.91

0.77

 Third interpretation

93

0.73–0.99

0.12

Reader 3

 Initial interpretation

68

0.44–0.86

 

 Second interpretation

61

0.37–0.81

0.66

 Third interpretation

93

0.73–0.99

0.017

Initial interpretation was based on conventional MR imaging findings including the morphologic features, signal intensity and contrast enhancement patterns of the lesion. Second interpretation was based on conventional MR imaging findings and ADC value of the lesion. Third interpretation was based on conventional MR imaging findings, ADC value and SUVmax of the lesion. The change in each reader’s accuracy between the initial and second interpretations and between the second and third interpretations was evaluated for statistical significance by using the pairwise comparison of ROC curves

Fig. 3

ROC curves for one observer (reader 2) based on the probability of a PCNSL assignment. On each graph the blue line represents the ROC curve generated from the first interpretation session. It was based on morphologic features, the SI, and the lesional contrast enhancement pattern on conventional MRI scans. The dashed brown line represents the ROC curve generated from the second interpretation session; it is based on conventional MRI findings and the lesional ADC value. The orange line represents the ROC curve generated from the third interpretation session; it is based on conventional MRI findings, the ADC value, and SUVmax of the lesion. Note that accuracy (expressed as the area under the ROC curve) of the third session that included SUVmax of the lesion (orange line) is improved compared to the first and second interpretation sessions (blue and dashed brown lines)

The addition of SUVmax at the third interpretation session resulted in higher correct tumor type identification by all 3 readers (Table 1). The mean accuracy for tumor differentiation was 75% (range 61–89%) at the second and 95% (range 93–100%) at the third interpretation. One reader achieved significantly higher accuracy at the third compared to the second interpretation (p = 0.017).

The κ values indicating the confidence level among the 3 readers for image interpretation are shown in Table 2. The addition of ADC for the second interpretation resulted in lower interobserver agreement between all pairs of readers. The addition of SUVmax at the third interpretation session produced higher interobserver agreement between all pairs of readers (Fig. 4; Table 2).
Table 2

Interobserver agreement of assessment for tumor discrimination with a five-point confidence scale

Comparison

Initial interpretation

Second interpretation

Third interpretation

R1 versus R2

0.61

0.46

0.64

R1 versus R3

0.37

0.19

0.62

R2 versus R3

0.44

0.40

0.60

Data are κ value

R1 reader 1, R2 reader 2, R3 reader 3

Fig. 4

A 75-year-old woman with PCNSL. a T2-weighted image demonstrating a heterogeneous hyperintense lesion (arrows) in the right basal ganglia and surrounding hyperintensity areas. b Contrast-enhanced T1-weighted image showing a ring-like enhanced lesion in the right basal ganglia. After evaluating the morphologic features, signal intensity, and contrast enhancement patterns of the lesion on conventional MRI, 2 of the 3 readers scored the lesion as probably GBM; the third recorded the finding as equivocal. c On the ADC map, the enhancing solid lesion is nearly isointense relative to the normal brain (arrow). The ADC value of the enhancing solid portion is 0.8 × 10−3mm2/s. After evaluating conventional MRI scans supplemented with the ADC value, 2 of 3 readers submitted a reading of probably GBM, the third considered the finding equivocal. d On FDG-PET images, the enhancing solid lesion exhibited a higher uptake than the normal cerebral cortices. SUVmax of the lesion is 21.1. After evaluating conventional MRI scans, ADC, and SUVmax, all 3 observers submitted a reading of probable PCNSL

Discussion

Overall, adding the tumor ADC value to conventional MRI studies resulted in slightly higher accuracy and lower interobserver agreement with respect to tumor differentiation. In one observer the addition of the ADC value to conventional MRI scans lowered accuracy. This finding suggests that adding the tumor ADC to conventional MRI studies fails to offer a sufficient advantage over the use of morphologic features, SI, and lesional contrast enhancement patterns in the differentiation between PCNSL and GBM. Earlier studies [9, 10, 11, 12] indicated that there are various ranges in the ADC value for PCNSL and GBM. The mean ADC values ranged from 0.63 to 0.87 × 10−3mm2/s for PCNSL and from 0.90 to 1.13 × 10−3mm2/s for GBM. The mean ADC values of PCNSL and GBM in our patients were similar to previously reported data, although the mean ADC values of PCNSL and GBM in our study were not significantly different.

On the other hand, the addition of the tumor SUVmax at the third interpretation session resulted in higher accuracy and interobserver agreement compared to assessments performed on conventional MRI supplemented with the ADC value. Overall, the addition of SUVmax improved reader accuracy by as much as 20%. More importantly, the reader with the highest accuracy (89%) at the second interpretation achieved improved accuracy after assessing the SUVmax data. This suggests that adding SUVmax to MRI provides useful additional information for the discrimination between PCNSL and GBM.

Although FDG-PET has been reported as a useful tool for differentiating PCNSL from high-grade glioma [13], the value of adding FDG-PET to conventional MRI for the differentiation of PCNSL and GBM has not been described. Kosaka et al. [13] identified SUVmax on FDG-PET as the most important parameter for distinguishing lymphomas from other brain tumors. They used a SUVmax of 15 as a cutoff for diagnosing CNS lymphoma and only one high-grade glioma (SUVmax, 18.8) returned a false-positive result. Based on their findings we used SUVmax on FDG-PET images in our study; the optimal cutoff point for SUVmax was 12. The accuracy of SUVmax alone for distinguishing PCNSL from GBM (cutoff point 12) was 86%; sensitivity was 100% and specificity was 71.4%. SUV measurements of brain tumors may be influenced by a wide variety of factors such as the plasma glucose level, steroid treatment, time after tracer injection, and previous irradiation [16, 17]. We carefully excluded patients with diabetes mellitus, steroid treatment, and/or previous irradiation.

As did others [4, 5, 6, 7], we document here that conventional MRI studies are valuable for distinguishing PCNSL from GBM. Although difficult cases with similar MRI findings, i.e., GBM with relatively homogeneous enhancement and PCNSL with a solitary mass and heterogeneous enhancement, were included in our study, approximately 75% of lesions could be differentiated on conventional MRI alone. Based on our results, we recommend that FDG-PET be added in cases where it is difficult to differentiate between PCNSL and GBM on conventional MRI scans and ADC maps.

Our study has some limitations. First, we used two different MR units at 1.5T and 3T. Although the absolute ADC values of the brain may vary at 1.5 and 3.0T, the difference of ADC value between 1.5-T and 3.0-T imagers from the same vendor was only 3–5% [18]. We measured ADC values of normal gray matter and white matter in the 21 patients performed by 1.5 or 3.0 T. The difference of ADC values of normal gray matter and white matter between 1.5- and 3.0-T imagers was within 5 and 4%, respectively. Therefore, we think that the variability in the absolute ADC values between these units in our study was an acceptable range. Second, our study design may have imposed bias. The observer performance study included 3 interpretation sessions per case: at the first session only conventional MRI were read, at the second session the ADC value was added to conventional MRI, and at the third, conventional MRI, the ADC value, and SUVmax were available for interpretation. As the ADC value was added at the second session, we were unable to evaluate the influence of only the addition of SUVmax to conventional MRI. In routine clinical practice, diffusion-weighted imaging including the ADC is widely used for assessing various pathologic conditions. Therefore, our second interpretation session that included the ADC value could be considered to reflect the interpretation of routine conventional MRI. Third, there is a difference between observer performance studies and the clinical environment. As we provided our observers with information on the two tumor types, their performance may have been affected. Because there were only two types, the answer to 50% of the questions would have been correct from the outset. Fourth, our study was retrospective and the study population was relatively small.

In conclusion, in a selected group of PCNSL and GBM with similar MRI findings, the addition of quantitative FDG-PET to MRI studies may improve the radiologists’ ability to differentiate between PCNSL and GBM.

Conflict of interest

The authors declare that they have no conflict of interest.

Copyright information

© The Japanese Society of Nuclear Medicine 2011

Authors and Affiliations

  • Keishi Makino
    • 1
  • Toshinori Hirai
    • 2
  • Hideo Nakamura
    • 1
  • Ryuji Murakami
    • 3
  • Mika Kitajima
    • 2
  • Yoshinori Shigematsu
    • 2
  • Rumi Nakashima
    • 4
  • Shinya Shiraishi
    • 2
  • Hiroyuki Uetani
    • 2
  • Koya Iwashita
    • 2
  • Masuma Akter
    • 2
  • Yasuyuki Yamashita
    • 2
  • Jun-ichi Kuratsu
    • 1
  1. 1.Department of Neurosurgery, Graduate School of Medical SciencesKumamoto UniversityKumamotoJapan
  2. 2.Department of Diagnostic Radiology, Graduate School of Medical SciencesKumamoto UniversityKumamotoJapan
  3. 3.Department of Radiation Oncology, Graduate School of Medical SciencesKumamoto UniversityKumamotoJapan
  4. 4.Japanese Red Cross Kumamoto Health Care CenterKumamotoJapan

Personalised recommendations