The incidence of non-Hodgkin’s lymphoma (NHL), especially extranodal lymphoma, has increased during the last several decades [
The importance of personalized precision medicine has been highlighted recently. In PGIL-DLBCL, considering the postoperative complications of traditional surgical resection [
The use of 18F-fluorodeoxyglucose (18F-FDG) positron emission tomography (PET)/computed tomography (CT) in Hodgkin’s lymphoma (HL) and aggressive NHL has been widely approved [
Intratumor heterogeneity, which correlates with tumor aggressiveness and poor prognosis, has been increasingly noted [
In the present study, we aimed to explore whether PET/CT TA was useful in predicting interim response in PGIL-DLBCL patients treated with chemotherapy and to compare the predictive values of texture features with those of the maximum standard uptake value (SUVmax) and MTV. We also aimed to obtain a prediction probability using texture features, clinical characteristics, and traditional PET semiquantitative features.
This retrospective study was approved by the local ethics committee, and the requirement for informed consent was waived. From June 2013 to March 2019, 60 patients with newly diagnosed PGIL-DLBCL were retrospectively reviewed. The inclusion criteria were as follows: (1) a diagnosis of PGIL-DLBCL confirmed by biopsy, (2) 18F-FDG PET/CT scan before treatment, and (3) an interval between the PET/CT scan and biopsy of less than 1 month. The exclusion criteria were as follows: (1) loss of follow-up (
Flowchart of patient inclusion and exclusion. The flowchart shows information about the inclusion and exclusion criteria that were used to ultimately include 30 patients with PGIL-DLBCL, including 20 in the CR group and 10 in the non-CR group.
Clinical and pathological information, including the involved sites in the GI tract, Lugano stage, international prognostic index (IPI), histological subtypes (germinal center B-cell-like (GCB) and non-GCB), and Ki67, was collected retrospectively from inpatient medical records and histologic reports.
The treatment plan for all eligible patients was 6–8 cycles of chemotherapy. A PET/CT scan was performed after 3-4 cycles of chemotherapy (29 after 4 cycles, 1 after 3 cycles) to evaluate the interim response. The interim response was assessed according to the PET-CT-based Lugano response criteria [
18F-FDG PET/CT scans were performed with a 16-row hybrid PET/CT scanner (Gemini GXL16, Philips Medical System, Cleveland, Ohio, USA). The serum glucose levels of all patients were confirmed to be less than 11.1 mmol/L after fasting for at least 6 hours. Then, 5.2 MBq (±10%) per kilogram of body weight of 18F-FDG was injected intravenously 50–90 minutes before PET/CT scanning. All patients were encouraged to drink 600–1000 ml of water 5 minutes before scanning to achieve gastric distension and were scanned in the supine position with arms elevated above the head and breathing at rest. For each patient, an unenhanced CT from the skull base to the upper thigh was performed for anatomic information and attenuation correction (CT scanning parameters: 50 mA, 120 kV, 5 mm section thickness, 5 mm increment, and a pitch of 0.813). The CT images were reconstructed to a 512 × 512 matrix. A 3-dimensional PET scan of the same region was subsequently obtained without any change in position. The emission data were acquired for 70 seconds per bed position, and a total of 8-9 bed positions were performed. The PET images were reconstructed in a 144 × 144 matrix with a voxel size of 4 mm × 4 mm × 4 mm and a slice thickness of 4 mm by a line-of-response algorithm using Syntegra software (Philips Corp., Amsterdam, Netherlands).
All PET/CT images were retrospectively reviewed by a radiologist (Y. S., with 9 years of experience in oncologic PET/CT) and confirmed by another radiologist (C. J., with 6 years of experience in oncologic PET/CT). Both radiologists had no knowledge of the results of the interim response assessment. Since we did not aim to explore the diagnostic value of PET/CT in DLBCL-PGIL, the tumor location was not blinded. The PET/CT images were transferred to the MedEx workstation (Beijing, China) to measure the SUVmax and MTV. The SUVmax and MTV were automatically generated by the MedEx workstation after each tumor was enclosed in a cropping sphere, and the MTV was defined as the volume of voxels with SUVs higher than the threshold of 41% × SUVmax.
The PET and CT images were uploaded to in-house software (Image Analyzer 2.0, China), and TA was performed separately on PET and CT images. In cases with multiple tumors in the GI tract, the tumor with the highest SUVmax was chosen for analysis.
In the PET images, regions of interest (ROIs) were manually drawn slice by slice to cover the entire volume of the tumors. The GI lumen and adjacent lesions (such as involved lymph node or liver tissue) were carefully avoided. The following first- and second-order texture features were derived from the PET images, including (1) first-order features: mean, standard deviation (SD), max-frequency, mode, minimum, maximum, cumulative percentiles (the 5th, 10th, 25th, 50th, 75th, and 90th percentiles), skewness, kurtosis, entropy, volume, and max-diameter and (2) local textural features of the grey-level co-occurrence matrix (GLCM): entropyGLCM, energyGLCM, inertiaGLCM, and varianceGLCM.
In each CT image, an ROI was manually drawn along the margin of the tumor on the section that depicted the largest area of the lesion, with artefacts and the gastrointestinal lumen carefully avoided. The attenuation value of each pixel within the ROIs was automatically read and analyzed by the software, and the following texture features were generated from CT images: mean, SD, max-frequency, mode, maximum, minimum, skewness, kurtosis, entropy, max-diameter, entropyGLCM, energyGLCM, inertiaGLCM, and varianceGLCM.
The Shapiro–Wilk normality test was applied to evaluate the distribution characteristics of the SUVmax, MTV, PET texture parameters, and CT texture parameters. The differences in the CR rate in patients with different clinicopathological characteristics were compared by Fisher’s exact test. Feature selection was processed by two steps: (1) univariate filtering was performed on all of the texture features using the Mann–Whitney
A total of 30 patients were ultimately included in our study cohort (11 males, 19 females; age range, 31–79 years; median age, 56 years; interquartile range, 47–63 years).
Among the 30 enrolled patients, 25 were treated with the R–CHOP protocol, while 5 were treated with other protocols that included rituximab. Two patients underwent PET/CT response assessments after 4 cycles of chemotherapy and then dropped out of the treatment plan (one died from severe interstitial pneumonia, and one turned to traditional Chinese medical therapy). Since the withdrawals were not expected at the time when they accepted the PET/CT response assessments, they were still considered to be “interim responses.”
In the PET/CT interim response assessment, 20 patients achieved CR (three with Deauville score 1, eight with Deauville score 2, and nine with Deauville score 3), while 10 patients did not achieve CR (three with Deauville score 4 and seven with Deauville score 5). The patients’ clinicopathological characteristics are presented in Table
Clinicopathological characteristics.
Characteristics | Number of patients ( |
---|---|
Gender | |
Male | 11 |
Female | 19 |
Age (years) | |
≤60 | 21 |
>60 | 9 |
Lugano stage | |
Stage I | 10 |
Stage II | 9 |
Stage IV | 11 |
Number of lesion(s) in GI tract | |
One | 26 |
Two or more than two | 4 |
Involved sites in GI tract | |
Fundus of stomach | 1 |
Body of stomach | 14 |
Antrum of stomach | 11 |
Duodenum | 4 |
Jejunum or ileum | 3 |
Ileocecal junction | 4 |
Colon | 4 |
Histological subtype | |
GCB-DLBCL | 13 |
Non-GCB-DLBCL | 17 |
IPI score | |
0 | 9 |
1 | 9 |
2 | 3 |
3 | 7 |
4 | 2 |
5 | 0 |
Interim response evaluation | |
CR | 20 |
Non-CR | 10 |
GI: gastrointestinal; GCB-DLBCL: germinal center B-cell-like diffused large B-cell lymphoma; IPI: international prognostic index; CR: complete remission.
The CR rates of different groups of stages, IPI scores, histological subtypes, involved sites, and Ki67 are shown in Table
CR rates in patients with different clinicopathological characteristics.
Characteristic | Number of CR | Number of non-CR | CR rate (%) | |
---|---|---|---|---|
Lugano stage | ||||
Stage I | 9 | 1 | 90.0 | 0.062 |
Stage II and IV | 11 | 9 | 50.5 | |
IPI score | ||||
0–2 | 16 | 5 | 76.2 | 0.104 |
3–5 | 4 | 5 | 44.4 | |
Histological subtype | ||||
GCB | 9 | 4 | 69.2 | 0.554 |
Non-GCB | 11 | 6 | 64.7 | |
Intestinal involvement | ||||
Involved | 6 | 7 | 46.2 | 0.045 |
Not involved | 14 | 3 | 82.4 | |
Ki67 | ||||
<80% | 9 | 3 | 75.0 | 0.350 |
≥80% | 11 | 7 | 61.1 |
Some texture features did not have a normal distribution. The detailed results of the normality test are shown in Supplemental Table
In the first step of feature selection, a total of 17 PET texture features and 24 CT texture features were found to be of no significant differences between the CR and non-CR groups and were eliminated. The detailed results of the Mann–Whitney
Among the remaining PET texture features, the mean, SD, max-frequency, 50th percentile, 75th percentile, 90th percentile, maximum, entropy, volume, max-diameter, entropyGLCM10, and entropyGLCM12 were significantly lower in the CR group, while the energyGLCM10, energyGLCM11, energyGLCM12, and energyGLCM13 were significantly higher in the CR group. The remaining CT texture features included the max-frequency and max-diameter, which were significantly lower in the CR group. The SUVmax and MTV were also significantly lower in the CR group (Table
Differences between the CR group and non-CR group.
Parameter | Median (interquartile range) | ||
---|---|---|---|
CR group | Non-CR group | ||
SUVmax | 15.15 (8.73–21.75) | 23.95 (21.45–29.03) | 0.001 |
MTV(cm3) | 17.80 (11.70–53.08) | 145.10 (66.63–613.90) | 0.009 |
PET texture features | |||
Mean | 3983.04 (3024.75–7348.02) | 7348.96 (5402.80–10440.45) | 0.028 |
SD | 1781.05 (1042.33–2554.85) | 2878.15 (1933.98–5193.57) | 0.044 |
Max-frequency | 3.00 (2.00–4.00) | 5.00 (3.00–7.25) | 0.019 |
Maximum | 9324.50 (6768.75–16080.00) | 17682.00 (11709.25–24430.00) | 0.031 |
50th percentile | 3705.50 (2620.50–6841.25) | 6699.00 (4841.00–9395.25) | 0.028 |
75th percentile | 4789.00 (3903.75–9199.00) | 8963.50 (6721.50–13698.00) | 0.035 |
90th percentile | 6081.00 (5408.50–10572.75) | 10257.00 (8360.00–16868.50) | 0.039 |
Entropy | 6.3777 (5.4630–7.0874) | 7.7240 (7.0291–8.5527) | 0.007 |
Volume (mm3) | 41504 (15920–85760) | 226272 (83760–683088) | 0.006 |
Max-diameter (mm) | 48.44 (36.50–75.96) | 85.62 (68.95–142.09) | 0.011 |
EntropyGLCM10 | 9.02 (7.16–9.76) | 11.45 (8.51–12.56) | 0.015 |
EntropyGLCM12 | 9.07 (7.36–9.84) | 11.38 (8.70–12.56) | 0.011 |
EnergyGLCM10 | 0.001880 (0.000961–0.005434) | 0.000363 (0.000172–0.002750) | 0.049 |
EnergyGLCM11 | 0.001550 (0.000863–0.004614) | 0.000364 (0.000167–0.002392) | 0.039 |
EnergyGLCM12 | 0.001757 (0.000951–0.005436) | 0.000388 (0.000172–0.002414) | 0.035 |
EnergyGLCM13 | 0.001544 (0.000821–0.004535) | 0.000338 (0.000167–0.002186) | 0.039 |
CT texture features | |||
Max-frequency | 31.50 (23.75–81.00) | 99.50 (46.00–276.25) | 0.011 |
Max-diameter (mm) | 52.05 (40.58–89.65) | 75.15 (64.88–124.18) | 0.024 |
In the second step of feature selection, the remaining features were categorized and selected as follows. (a) Among the features describing FDG uptake intensity, including the mean, 50th percentile, 75th percentile, 90th percentile, and maximum (
The predictive values of the SUVmax, MTV, and selected texture features for interim response were evaluated by ROC analyses, and the results are displayed in Table
ROC analysis of SUVmax, MTV, and texture features.
Parameter | Cutoff | Sensitivity | Specificity | Accuracy | AUC | |
---|---|---|---|---|---|---|
SUVmax | 18.6 | 1.00 | 0.75 | 0.83 | 0.850 | <0.001 |
MTV (cm3) | 49.7 | 0.90 | 0.75 | 0.80 | 0.790 | 0.006 |
PET texture features | ||||||
50th percentile | 4139 | 0.90 | 0.70 | 0.77 | 0.750 | 0.012 |
Entropy | 7.13 | 0.80 | 0.80 | 0.80 | 0.800 | <0.001 |
Volume (mm3) | 85824 | 0.80 | 0.80 | 0.80 | 0.805 | <0.001 |
CT texture features | ||||||
Max-frequency | 44 | 0.90 | 0.70 | 0.77 | 0.783 | 0.001 |
ROC analysis of parameters with AUCs ≥0.800 and the prediction probability. (a) ROC analysis of the SUVmax, volume, and entropy. The AUCs were 0.850, 0.805, and 0.800, respectively. (b) ROC analysis of the prediction probability generated from the combination of the SUVmax, entropy, volume, and intestinal involvement. The AUC was 0.915.
Intestinal involvement and the SUVmax, volume, and entropy were selected to be included to generate a prediction probability. The Hosmer–Lemeshow test showed a chi-square value of 9.727 and a
The median MTV and median volume were 45.20 cm3 (interquartile range = 13.88–127.45 cm3) and 59.87 cm3 (interquartile range = 20.66–260.91 cm3), respectively. There was a significant difference between the MTV and volume according to the Wilcoxon signed rank test (
The SUVmax, MTV, some PET texture features (mean, SD, maximum, 90th percentile, 75th percentile, 50th percentile, max-frequency, volume, max-diameter, entropy, inertiaGLCM11, inertiaGLCM12, inertiaGLCM13, varianceGLCM10, varianceGLCM11, varianceGLCM12, and varianceGLCM13), and some CT features (max-frequency, max-diameter, entropyGLCM13, inertiaGLCM10, inertiaGLCM11, inertiaGLCM12, and inertiaGLCM13) showed excellent interobserver agreement. Other PET and CT texture features showed good-to-poor interobserver agreement. The detailed ICCs are shown in Supplemental Table
The present study explored the use of the SUVmax, the MTV, PET/CT texture features, and clinicopathological characteristics in predicting the interim treatment response of PGIL-DLBCL. We found that the SUVmax, the MTV, some texture features, and the tumor location were useful parameters in interim response prediction. Moreover, employing a combination of the pretreatment SUVmax, texture features (entropy and volume), and intestinal involvement further improved the predictive value in PGIL-DLBCL patients.
Previous studies have demonstrated that a high SUV is associated with a poor prognosis [
PET TA provides information about the intratumor heterogeneity of FDG uptake noninvasively from routine images [
In the current study, some first-order texture features, including the mean, 50th percentile, 75th percentile, 90th percentile, and maximum, were found to be significantly higher in the non-CR group, with AUCs ranging from 0.735 to 0.750. These features reflect the degrees of FDG uptake of the pixels and provide detailed information on FDG distribution.
The energy of GLCM is calculated by the formula
Entropy quantitatively characterizes the intratumor heterogeneity. The more chaotically the intensities of the pixels are distributed, the higher the entropy [
In the present study, a high volume and high MTV were found to be predictors of non-CR. The volume and MTV are similar parameters that indicate the tumor burden but are measured by different methods (the volume was derived from the manually drawn ROI, while the MTV was generated automatically by a computer program based on a set threshold). Multiple studies have demonstrated that a high MTV is associated with an insufficient treatment response and a poor prognosis in lymphoma [
In addition, intestinal involvement was found to be a predictor of non-CR in the present study. Previous studies have reported poorer prognoses of intestinal lymphoma than gastric lymphoma [
The SUVmax, entropy, volume, and intestinal involvement were chosen and combined to generate a prediction probability. This combination characterized the tumors from different perspectives, namely, glucose metabolism, intratumor heterogeneity, tumor burden, and anatomical site. The prediction probability was demonstrated to be an excellent predictor of the interim response with an AUC higher than any single parameter (AUC = 0.915).
The interobserver ICCs were calculated to evaluate the reproducibility of the texture features. The SUVmax, MTV, mean, SD, maximum, higher percentiles (50th, 75th, and 90th), first-order entropy, volume, and max-diameter of PET images were found to be of excellent interobserver reproducibility, with ICCs ranging from 0.807 to 0.988. However, the first-order skewness and kurtosis had relatively low ICCs (0.515 and 0.430, respectively). These results accorded with those of a previous study [
The current study has several limitations. First, the present study was preliminary and retrospective. The study cohort was small, as it was limited by the low incidence of PGIL, the filtered histological subtype of DLBCL, and the exclusion of patients who did not accept consecutive chemotherapy and PET/CT scans. Some PET/CT scans were performed beyond the recommended interval between FDG administration and acquisition [
The preliminary study indicated that TA had potential for improving the value of pretreatment PET/CT in predicting the interim response in PGIL-DLBCL. However, prospective studies with increased sample sizes and validation analyses should be performed to confirm the present findings.
The data used to support the findings of this study are available from the corresponding author upon request.
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Yiwen Sun and Xiangmei Qiao contributed equally to this manuscript.
Table 1: normality test of SUV and texture features. Table 2: features without significant differences between the CR and non-CR groups in the Mann–Whitney