This retrospective study was approved by the hospital’s ethics committee (No. S2022-671-01), and the requirement for written informed consent was waived. Between January 2013 and July 2023, 69 pathologically proved PCNSL patients were retrieved from the PACS of our hospital. Four of them were excluded due to the usage of steroid treatment before MR scan or blurred MR images. Finally, 65 patients (101 lesions) were included in this study, among whom 40 were BCL-6 overexpression (23 males and 17 females, mean age, 58.11 ± 13.65 years) and 25 were BCL-6 underexpression (16 males and 9 females, mean age, 56.59 ± 12.40 years). Patients were included according to the following criteria: (1) pathologically confirmed DLBCL; (2) immunohistochemical staining results of BCL-6 (> 30% was defined as BCL-6 overexpression [BCL-6 (+)], ≤ 30% was defined as BCL-6 underexpression [BCL-6 (-)]); (3) patients were immunocompetent (determined by past history and laboratory tests including HIV, lymphocyte ratio, IgG, etc.); and (4) over 18 years of age. The exclusion criteria included: (1) incomplete MRI protocol or obvious artifacts in MR images; (2) MR scans after chemotherapy, radiotherapy, or steroid usage; (3) MR scans after tumor resection; and (4) incomplete histological results. All the clinical data, MRI protocol, and MR image artifacts were collected and assessed by A (6-year work experience), B (6-year work experience), and C (3-year work experience). The flowchart of patient inclusion was shown in Fig. 1. Risk score was assessed using three categories (low, intermediate, and high) based on the International Extranodal Lymphoma Study Group (IELSG) [19] and the Memorial Sloan-Kettering Cancer Center prognostic model (MSKCC) [20], respectively.

Flowchart of the study cohort. PCNSL = primary centeral nervous system lymphoma

MR images were acquired on a 3T MRI system (Discovery 750, GE Healthcare, Milwaukee, WI) with a 32-channel head coil. MR imaging protocol included axial fast spin echo T2 weighted imaging (T2WI, repetition time [TR]/echo time [TE] = 5642/93 msec, field of view [FOV] = 24 × 24 cm, matrix = 512 × 512, number of excitation [NEX] = 1.50), coronal T2 fluid-attenuated inversion recovery (T2FLAIR) (TR/TE/inversion time [TI] = 8527/162/2100 msec, FOV = 24 × 24 cm, matrix = 288 × 224, NEX = 1.00), axial diffusion-weighted imaging (DWI, TR/TE = 3000/65.5 msec, FOV = 24 × 24 cm, b = 0/1000 sec/mm2, matrix = 160 × 160, and NEX = 2.00), with a slice thickness of 5.0 mm and gap of 1.5 mm for all the axial and coronal images.

DICOM images were converted to the NIfTI format, and the image structure was segmented using ITK-SNAP [21]. The areas of interest were comprised of tumor solid components and necrosis or cystic changes, except the hemorrhage. Meanwhile, lesions with a minimum diameter of more than 5 mm were included so as to ensure a reliable measurement. Three dimension Segmentations were delineated on ADC, T2WI, and T2FLAIR, respectively, and intergroup consistency by intraclass correlation coefficient (ICC) was conducted. Regions of interest were drawn by A (6-year work experience) and B (6-year work experience). Tumor segmentation on different MRI sequences was demonstrated in Fig. 2.

An example of tumor segmentation. The red label represented the solid part of the tumor

Radiomics features were extracted by a customized version of Pyradiomics from Python (Version 3.8.8). A total of 2234 radiomics characteristics were extracted from each tumor segment in each set of images used in the machine learning models. The interpretation and calculation formulas, as well as all radiomics features, were presented on Pyradiomics website (https://pyradiomics.readthedocs.io/en/latest/features.html). The extracted features contained 14 shape features, 36 first-order histogram features, and 75 s-order features (also called texture features).

Filters such as the original image, 8 wavelet transforms, 5 log-sigma transforms, square, square root, logarithmic, exponential, gradient, and LBP-2D transform images were used to select the features.

Feature reduction was performed through the following steps.1) The extracted features with ICC ≥ 0.80 were selected. 2) Levene assay was used to analyze the normal distribution of test features. The correlation between features and BCL-6 expression was then assessed by t-test, and features with P < 0.05 were considered as potential predictors. 3) Least absolute shrinkage and selection operator (LASSO) regression was used to select features. 4) The final subset of features was filtered based on this number and the ranking of the feature weights.

Based on the 3 MRI sequences, 6 sequence groups (ADC, T2WI, T2FLAIR, ADC + T2WI, ADC + T2FLAIR, and ADC + T2WI + T2FLAIR) were selected. One sequence group (T2 + T2FLAIR) was deleted since no features were presented. Subsequently, five possible classifiers built on the machine learning tool kit of scikit-learn (https://scikit-learn.org) were used to adapt to these combinations. The following classifiers were used: logistic regression (LR), Naive Bayes (NB), support vector machine (SVM), K-nearest neighbor (KNN), and Multilayer Perception (MLP). The best features were combined with five different machine learning algorithms to filter the optimal model. Five-fold cross-validation was performed in the training set. Finally, the validity of 30 different models were evaluated. Default parameters were used for all machine learning models in this study.

Based on previous studies, [22–23] tumor morphology was selected and evaluated: (1) Tumor location was determined by the lesion center and divided into two groups: supratentorial or infratentorial, peripheral or median parenchyma (basal ganglia, thalamus, brainstem, and cerebellum). (2) The maximum and minimum diameters of the maximum layer of tumor were measured on T2WI. (3) ADC values, including ADCmean, ADCmax, ADCmin, ADC percentile, and ADC uniformity were calculated (4) Necrosis or cystic changes were defined as areas showing hyperintensity on T2WI.5) Edema (Table S1 in the Supplemental Material) and mass effect levels (Table S2 in the Supplemental Material). These tumor morphologic features were assessed by A. (6-year work experience), B (6-year work experience), and C (3-year work experience). Representative case images in BCL-6(+) and BCL-6(-) were shown in Fig. 3.

MRI in a 47-year-old male PCNSL patient with BCL-6(-). The tumor shows isointensity on axial T2WI (a), restriction on axial ADC map (b), and slight hyperintensity on coronal T2FLAIR (c). MRI in a 76-year-old female PCNSL patient with BCL-6(+). The tumor shows slight hyperintensity on axial T2WI (d), restriction on axial ADC map (e), and isointensity on coronal T2FLAIR (f)

The statistical product service solutions (version 26.0) were utilized to analyze statistical differences in clinical data and MR imaging features of lesion between BCL-6 (+) and BCL-6 (-) groups. Continuous variables were assessed by the t-test, while categorical variables were assessed by the nonparametric test. P value < 0.05 were considered statistically significant. The inter group consistency was tested by ICC, and ICC ≥ 0.80 was considered to be a strong correlation. Sensitivity, specificity, accuracy, F1-score, and area under curve (AUC) for each machine learning model were calculated.