Improving the Classification of PCNSL and Brain Metastases by Developing a Machine Learning Model Based on (18)F-FDG PET

Background: The characteristic magnetic resonance imaging (MRI) and the positron emission tomography (PET) findings of PCNSL often overlap with other intracranial tumors, making definitive diagnosis challenging. PCNSL typically shows iso-hypointense to grey matter on T2-weighted imaging. However, a particular part of PCNSL can demonstrate T2-weighted hyperintensity as other intracranial tumors. Moreover, normal high uptake of FDG in the basal ganglia, thalamus, and grey matter can mask underlying PCNSL in (18)F-FDG PET. In order to promote the efficiency of diagnosis, the MRI-based or PET/CT-based radiomics models combining histograms with texture features in diagnosing glioma and brain metastases have been widely established. However, the diagnosing model for PCNSL has not been widely reported. The study was designed to investigate a machine-learning (ML) model based on multiple parameters of 2-deoxy-2-[18F]-floor-D-glucose ((18)F-FDG) PET for differential diagnosis of PCNSL and metastases in the brain. Methods: Patients who underwent an (18)F-FDG PET scan with untreated PCNSL or metastases in the brain were included between May 2016 and May 2022. A total of 126 lesions from 51 patients (43 patients with untreated brain metastases and eight patients with untreated PCNSL), including 14 lesions of PCNSL, and 112 metastatic lesions in the brain, met the inclusion criteria. PCNSL or brain metastasis was confirmed after pathology or clinical history. Principal component analysis (PCA) was used to decompose the datasets. Logistic regression (LR), support vector machine (SVM), and random forest classification (RFC) models were trained by two different groups of datasets, the group of multi-class features and the group of density features, respectively. The model with the highest mean precision score was selected. The testing sets and original data were used to examine the efficacy of models separately by using the weighted average F1 score and area under the curve (AUC) of the receiver operating characteristic curve (ROC). Results: The multi-class features-based RFC and SVM models reached identical weighted-average F1 scores in the testing set, and the score was 0.98. The AUCs of RFC and SVM models calculated from the testing set were 1.00 equally. Evaluated by the original dataset, the RFC model based on multi-class features performs better than the SVM model, whose weighted-average F1 scores of the RFC model calculated from the original data were 0.85 with an AUC of 0.93. Conclusions: The ML based on multi-class features of (18)F-FDG PET exhibited the potential to distinguish PCNSL from brain metastases. The RFC models based on multi-class features provided comparatively high efficiency in our study.