2.1 Patients
In this retrospective study, the patient cohort from a hospital consisted of 88 patients who experienced recurrence between September 2010 and December 2023. This study passed the medical ethics review of the Branch of Medical Ethics Committee, and as it was a retrospective study, the requirement for informed consent was waived. All workflows were in line with standard regulations and guidelines. The recurrence criteria in this study consisted of radiological and clinical standards. Radiological criteria rely on imaging examinations to determine tumor recurrence, typically manifesting as new abnormal signals or enhanced areas at the primary tumor site. Clinical criteria assess whether patients develop new neurological symptoms or experience disease progression.
Two common radiotherapy modalities are often used in the clinical treatment of glioblastoma in hospitals: Gamma Knife Radiosurgery (GKRS) and conventional linear accelerator radiotherapy. However, no definitive conclusions have been drawn about the difference in the impact of the two treatments on the survival of glioblastoma patients. Given the extensive use of GKRS in our collaborating hospital and the results of our previous study [21] showing that it can significantly prolong OS and PFS, especially in patients with small recurrent lesions, short intervals, high KPS scores and multiple GKRS treatments, coupled with its lower radiation-induced damage to healthy tissue, we adopted a controlled variable approach and selected data from recurrent glioblastoma patients treated with GKRS for our study.
The inclusion criteria were as follows: (a) histopathologically confirmed, recurrent, grade I-IV glioma; (b) clinical variables and corresponding follow-up records; and (c) treatment with gamma knife radiosurgery (GKRS).
Patients were excluded if (a) MRI sequences were absent when gliomas recurred (n = 8). Finally, a total of 80 patients were randomly divided into two groups at a ratio of 3:1. The complete inclusion and exclusion criteria used in this study are shown in Fig. 1.
Position for Fig. 1. Flowchart of the inclusion and exclusion criteria.
The clinical characteristics of the patients, including demographics (sex, age), pre-GK treatment specifics (initial surgical time, multiple craniotomies, surgery-to-GKRS interval, adjuvant treatments after initial surgery, KPS scale), GKRS treatment parameters (number of targets, volume of PTV, maximum dose, marginal dose, central does, peritumoral does, multiple GKRS, concurrent/adjuvant chemotherapy), recurrence interval, OS, number of GKRS procedures, clinical notes, radiological images, and reports, and telephone follow-up, were recorded.
All T1-weighted FLAIR images were obtained in the routine clinical workup with two MR scanners from 1.5T Signal HDxt and 3.0T Discovery MR 750 W (GE Healthcare, Fairfield, Connecticut, USA). The parameters were repetition time (TR)/echo time (TE), 2317 ms/11 ms; inversion time (TI), 860 ms; 3-mm-thick sections; field of view (FOV), 225 × 225 mm2 for the GE Signal; TR/TE, 4.044 ms/1.834 ms; and 1.5-mm slice thickness for the GE Discovery MR 750 W.
2.2 Image feature extraction
One radiologist initially delineated the treatment target area, which was subsequently reviewed and confirmed by two senior doctors. The voxels were resampled to 1 mm × 1 mm × 1 mm using trilinear interpolation [22]. The detailed parameter settings are listed in Supplementary Table S1. The flowchart of the survival analysis in this study is shown in Fig. 2. A total of 1300 image features were extracted from the region of interest (ROI) on T1-weighted FLAIR images with an open-source Python tool named PyRadiomics 2.2.2. The feature pool contains I) first-order features, Ⅱ) shape and size features, and Ⅲ) textural features. Supplementary Table S2 shows the details of the radiomic features.
Position for Fig. 2. Flowchart of the radiomics model for survival analysis.
The reproducibility and repeatability of radiomic features are critical problems introduced by image acquisition, preprocessing, and feature extraction. A three-step procedure reduced feature dimensionality. First, features dependent on the imaging scanner used were removed using Kruskal‒Wallis tests (p > 0.05). Second, correlated features (correlation > 0.9) were eliminated by removing one. Third, LASSO regression with cross-validation was repeated 300 times with different seeds to obtain stable feature selection. The final features met two criteria: 1) were selected in most repetitions, and 2) had a C-index > 0.7. In each repetition, 10-fold cross-validation was used to determine λ to minimize error, and LASSO regression was used to select features by setting irrelevant coefficients to zero. The selected features formed the final predictive model, which maintained its predictive accuracy. The third step intermediate results are provided in Supplementary Fig. S1.
A radiomics score was then computed for each patient by a linear combination of selected features weighted by their respective coefficients. The formula for the combined radiomics features (referred to as the radio-score) is as follows:
$$\sum _{i}^{m}{x}_{i}{w}_{i}= Radio\_score \left(1\right)$$
where \({w}_{i}\) is the weight of the retained features \({x}_{i}\).