Patient selection
This study involved 30 HGG patients who underwent tumor resection at Shandong Cancer Hospital or the First Affiliated Hospital of Shandong First Medical University between June 2017 and July 2019. The inclusion criteria were as follows: (1) age ≥ 18years; (2) preoperative Karnofsky Performance Status (KPS) ≥ 80; (3) tumor located in nonfunctional area and successful supra total resection; (4) tumor removal achieved with resection margins that included the neighboring normal tissue (between 2cm and 3cm away from the tumor border); and (5) no tumor progression observed during follow-up (at least ≥ 6months). The exclusion criteria included: (1) a medical history of brain chemoradiotherapy (CRT) and (2) multicentric or multifocal cerebral lesions. All tumors were graded according to the 2016 World Health Organization (WHO) classification [11]. This study was approved by the institutional review board. All patients provided written informed consent to participate in the trial.
Preoperative MRI Acquisition
Before any treatment, gadolinium-enhanced MRI examination was performed in all patients. MRI scanning was acquired using a 3T MRI scanner (Philips Achieva 3T). The scanning protocols were obtained with the following parameters. T1-weighted imaging: echo time (TE) = 10ms, repetition time (TR) = 495ms, slice thickness/gap = 3mm/0mm, number of signal averaged (NSA) = 1, field of view (FOV) = 260mm × 260mm, matrix = 256 × 256. T2-weighted imaging: TE = 110ms, TR = 13312ms, slice thickness/gap = 3mm/0mm, NSA = 1, FOV = 260mm × 260mm, matrix = 416 × 416. T2-fluid attenuated inversion recovery (T2-FLAIR): TE = 120ms, TR = 11000ms, slice thickness/gap=3mm/0mm, NSA = 1, FOV = 260mm × 260mm, matrix = 320 × 320. To accurately match the MRI and the tissue specimen, the orbitomeatal line (OML) was perpendicular to the scanning table.
Preoperative MRI was evaluated by a senior radiologist and the data (T2-FLAIR and T1-weighted sequence) were used to determine the PTBE volume (VPTBE), Vtumor and tumor localization. Based on the spatial relationship between the tumor and the subventricular zone (SVZ) and cortex, the tumor location was classified as follows: Type I, tumor contacting SVZ; Type II, tumor involving cortex; Type III, tumor neither contacting SVZ nor infiltrating cortex (Figure 1).
Surgical specimen processing
After resection, the plane of the OML was marked on the specimen. Subsequently, the surgical specimen was oriented according to the in vivo geometry, marked with different colors to indicate the original orientation of the specimen in the brain and fixed in 10% formalin (≥24h). The dimensions of the tumor samples, both before and after fixation were documented to determine the reduction in size due to fixation (Supplementary Table S1). Then, the plane of the OML was perpendicular to the table; the specimen was cut using a tissue slicer from the cranial to the caudal side in approximately 3mm thick slices, which ensured that each specimen slice could match the MRI slice. The slices were contiguous, and their individual thickness was measured with a ruler. Finally, whole-mount paraffin sections were made and cut into 5µm sections per slice, which were stained with hematoxylin and eosin (H&E) (Figures 2a-e). In addition, each patient underwent molecular testing, and the methods used for analyzing the methylation status of the MGMT promoter and determining the mutational status of IDH by DNA pyrosequencing have been described previously [12,13]. Deletions of chromosomes 1p/19q were evaluated by fluorescence in situ hybridization analysis in tumor tissue sections [12,13].
ME analysis and measurement
The tumor-containing area and the PTBE area in the H&E sections were microscopically outlined, scanned and recorded by TissueFAXS PLUS (TissueGnostics, Austria). Subsequently, the scan images were imported into Photoshop (Adobe Systems, USA) to identify the microscopic evidence by two experienced pathologists who were blinded to the clinical data. Invasive GCs were identified by means of their nuclear atypia and heteropyknotic staining [14] (Figure 2f). To measure the spatial distribution of invasive GCs, pathologic slices were used to generate three-dimensional (3D) graphics through 3D-DOCTOR (Able Software Corp, USA). First, the contours of individual H&E sections were digitized and recorded to generate a 3D surface of the reconstructed specimen. Second, the 3D specimens were correct for retraction through scaling parameters (Supplementary Table S1). Then, the corrected 3D image was registered to preoperative T1-weighted MRI using the outline of tumor to perform point-based registration. After the above steps, the ME distance and direction of the GCs relative to the primary tumor bulk were established. In the in-slice direction, the nearest (Euclidean) distance [15] from the edge of the tumor to the microscopic GCs surrounding brain tissue was measured by Photoshop (Figure 2g). In the through-slice direction, the number of slices from the invasive GCs to the lesion border was counted and multiplied by the slice thickness (×3mm). The ME of each slice is defined as the maximal distance of the ME. The ME of each patient was defined as the maximum ME across different slices (Supplementary Table S2).
Predictive model development and validation
Based on the multivariate linear regression (MVLR) analysis, we constructed a ME predictive model. To validate the feasibility and safety of this model, we prospectively collected another 30 HGG patients who underwent surgery at Shandong Cancer Hospital between September 2019 and January 2020. The following inclusion criteria were applied: (1) adult patients; (2) preoperative KPS ≥ 70; and (3) no medical condition that could interfere with oral medication intake. Patients who had a history of a previous cranial surgery, CRT, or contraindication for MRI were excluded. Patients were randomly split into two arms in a 1:1 ratio to receive guideline-based CRT (GroupNCCN) or model-based CRT (GroupModel).
All patients started treatment within 2 weeks after surgery and were treated with CRT, which delivered 60Gy in 30 fractions with continuous daily temozolomide (TMZ) (75mg/m2/d), followed by 6 cycles of adjuvant TMZ (150–200mg/m2 for 5/28days). Radiation treatment planning was performed with the Varian Eclipse Treatment Planning System. Target volumes were delineated according to co-registration of postoperative CT and MRI obtained with each patient in the treatment position. The management of GroupNCCN was designed according to the NCCN guidelines, Version 2.2019. GTV was defined by T1-weighted abnormality on the MRI, which consisted of all postoperative-enhanced MRI and the surgical cavity. The CTV was defined by GTV plus a margin of 2cm, adjusted to anatomical borders. The CTV was expanded by 3mm to create the respective planning target volume. For GroupModel, the CTV was defined as the GTV plus a margin, which was determined by our model. The rest of the treatment design was the same as that of GroupNCCN.
MRI was repeated before concurrent CRT, before the first cycle of adjuvant TMZ, and thereafter every 2-6weeks. Tumor progression and recurrence were identified by both the oncologist and the radiologist. Progression criteria were described by Macdonald et al. [16]. Recurrence patterns were defined as in-field if ≥ 80% of the lesion resided within the prescription 95% isodose line (D95) and marginal if 20-80% of the recurrence was inside the D95. For other cases, recurrences were defined as out-field [17]. The primary endpoint was the overall response rate (ORR) (the proportion of complete response and partial response). Secondary endpoints included progression-free survival (PFS) (the date from surgery to either first documented progression or death), and overall survival (OS) (the date from surgery to the date of death or the last follow-up).
Statistical analysis
For all analyses, we used SPSS 22.0 (IBM Armonk, NY, USA), and values for which P < 0.05 were considered statistically significant. Categorical variables were expressed as proportions and continuous variables were expressed as mean ± standard deviation. The difference between two groups was assessed with Student’s t-test or Chi-Squared test. When comparing more than two variables, we performed one-way analysis of variance. Post-hoc analysis was used to compare pairwise differences. Spearman’s rank correlation was performed to evaluate the relationship of the ME with the grade, Vtumor, location, VPTBE and molecular marker status. A MVLR model was created from variables with a P < 0.05 on correlation analysis, using stepwise regression. To assess the prediction efficiency of this model, calibration was evaluated using the R-square goodness-of-fit test, and discrimination was evaluated using receiver operating characteristic (ROC) curves with the corresponding area under the curve (AUC). The patient survival rates were determined by Kaplan-Meier curves and survival curves were analyzed by the log-rank test.