Deep Learning-based Segmentation and Discrimination of High-grade Gliomas and Solitary Metastatic Tumors

This study constructs a convolutional neural network (CNN) model to differentiate between high-grade gliomas and solitary metastases using peritumoral edema regions from preoperative magnetic resonance imaging (MRI) scans. We collect T1-enhanced sequences (T1CE) MRI image data from multiple scanners at two different medical centers and fine-tune the "Segment Anything Model" (SAM) using adapter technology on tumor datasets to automatically segment high-grade gliomas and solitary metastases for training the prediction model. We primarily conduct independent analysis of peritumoral edema regions for each case, evaluating the effectiveness of tumor differentiation using only the edema regions. Predictions for each patient are assessed using slice voting on the validation set. In five-fold cross-validation, our model achieves an average accuracy of 83.32%, sensitivity of 83.32%, and specificity of 83.32%. Additionally, on an external test set, our model achieves an average accuracy of 82.75%, sensitivity of 81.82%, and specificity of 83.33%. Our trained CNN model accurately distinguishes between high-grade gliomas and solitary metastases using peritumoral edema regions from preoperative MRI scans. This suggests that single-modality T1CE images are sufficient to differentiate between these two types of tumors, emphasizing the significant impact of peritumoral edema regions on tumor differentiation.