Related Experiment Video
Updated: May 12, 2026
Manual Segmentation of the Human Choroid Plexus Using Brain MRI
Published on: December 15, 2023
MixBranchNet: a task-adaptive network for glioma segmentation and genotype prediction by exploiting spatial-spectral
Yanduo Li1, Zhekai Chen1, Yitian Fan1
1Department of Electronic Science, Fujian Provincial Key Laboratory of Plasma and Magnetic Resonance, School of Electronic Science and Engineering, National Model Microelectronics College, Xiamen University, Xiamen, China.
Objective:
To develop a task-adaptive deep learning-based model, termed MixBranchNet, that leverages spatial-spectral correlations in chemical exchange saturation transfer (CEST) MRI for improved glioma segmentation and genotype prediction.
Methods:
MixBranchNet incorporates a dual-branch Mixing Block that extracts spatial and spectral features in parallel through convolutional and self-attention pathways. The model was trained on multi-offset CEST images with 41 frequency offsets from a 3.0 T MRI scanner. Segmentation and genotype prediction tasks were trained separately, with the segmentation output used as the input for the genotype prediction. Manual annotations from two board-certified neuroradiologists were used as the reference standard for segmentation, while genotype labels were obtained from surgical histopathology. Model performance was assessed using the Dice coefficient for segmentation, mean probabilities (Pmean), accuracy, sensitivity, specificity, F1-score, and AUC for genotype prediction. Five-fold cross-validation was performed within the development cohort using strict patient-level partitioning. A hold-out test set was defined prior to cross-validation and remained fully isolated from training, validation, and model selection procedures.
Results:
MixBranchNet achieved a Dice coefficient of 0.84 (95% CIs: 0.80-0.88) for segmentation, outperforming the full Z-spectrum-based fully convolutional network (FCN) (0.61, p < 0.001), the MedSAM model based on spatial information (0.80, p < 0.001), and the conventional CEST-specific U-Net (0.82, p = 0.003). For genotype prediction, MixBranchNet yielded a Pmean of 92.66% (95% CIs: 89.44% - 95.27%), and accuracy of 95.00% for IDH mutation, as well as a Pmean of 91.07% (95% CIs: 88.98% - 93.37%), and accuracy of 93.11% for MGMT promoter methylation. All results significantly exceeded the performance of conventional CEST quantification techniques and existing deep learning-based models developed for CEST analysis (p < 0.05).
Conclusion:
MixBranchNet establishes a methodological foundation for spatial-spectral deep learning in CEST MRI and demonstrates encouraging performance for glioma segmentation and genotype prediction within the current single-center cohort.
