CSF1R and macrophage infiltration: Integrated magnetic resonance imaging radiomics and deep learning-driven models for the preoperative assessment of glioma.

backgroundColony-stimulating factor-1 receptor (CSF1R) signaling is crucial for the ability of tumor-associated macrophages (TAMs) to establish an immunosuppressive tumor microenvironment (TME), highlighting the potential of CSF1R signaling as a therapeutic target. Noninvasive preoperative prediction of CSF1R levels in gliomas using magnetic resonance imaging (MRI) holds clinical potential for guiding immunotherapy.

methodsThis study enrolled 477 patients with glioma from three datasets. We retrospectively collected CSF1R staining data and paired MR images from center 1 between January 2020 and December 2022 (training cohort, n = 65) and extracted conventional radiomics (CR) and deep learning (DL) features, which were then integrated (CR+DL) to construct CSF1R prediction models using 12 classic machine learning classifiers, followed by five-fold cross-validation. We subsequently tested the model's performance using CSF1R staining data from prospective patients with glioblastoma (GBM) collected between April 2024 and September 2024 (internal test cohort, n = 38). External validation of the model, including its correlation with CSF1R gene expression, was performed using MRI-transcriptomic-prognostic paired data ( n = 101) from center 2, along with immune infiltration and survival analyses. Finally, the models were assessed from multiple perspectives using data from different cohorts at center 3: survival efficacy ( n = 255), correlation with macrophage immunohistochemical (IHC) staining in patients with GBM ( n = 16), and single-cell sequencing (scRNA-seq) ( n = 2, lesions = 4) data from patients with multifocal GBM.

resultsWe successfully developed CSF1R prediction models leveraging CR, DL, and CR+DL features. In the internal test cohort, the CR+DL model achieved the highest accuracy (0.76, support vector machine (SVM) classifier) and sensitivity (0.86, SVM classifier), whereas the DL model yielded the best specificity (0.88, random forest classifier) and the CR model yielded the best area under the curve (0.77, naive Bayes classifier). A significant correlation was identified between CSF1R prediction and CSF1R gene expression in patients with GBM ( P = 0.017). Further analysis of immune infiltration in samples from patients with GBM revealed higher immune scores and increased M2 macrophage infiltration in the groups with high CSF1R expression ( P = 0.0038). Finally, results from the IHC staining of macrophages ( P = 0.0003) and scRNA-seq data corroborated these findings.

conclusionWe successfully developed a CSF1R prediction model based on CR and DL features, elucidating its association with macrophage infiltration and its potential to guide preoperative immunotherapy strategies.