Clinical Report: DPEA-Net: A Lightweight 3D Convolutional Neural Network for Glioma Segmentation
Overview
DPEA-Net is a novel lightweight 3D convolutional neural network designed for accurate glioma segmentation from multiparametric MRI. It achieves high mean Dice scores while maintaining a low computational footprint.
Background
Accurate segmentation of glioma subregions is crucial for effective radiotherapy planning and patient management. Traditional manual segmentation methods are time-consuming and prone to variability.
Data Highlights
Segmentation Type
Mean Dice Score (BraTS 2019)
Mean Dice Score (BraTS 2020)
Whole Tumor (WT)
90.43%
89.96%
Tumor Core (TC)
85.56%
86.52%
Enhancing Tumor (ET)
81.89%
80.31%
Key Findings
DPEA-Net reduces parameters by 99% compared to 3D U-Net.
It enables adaptive multi-scale feature extraction to address tumor heterogeneity.
The model achieves sub-2-second inference on standard clinical hardware.
A 1.5-fold TC weighting strategy enhances segmentation of the tumor core.
Mean Dice scores for WT, TC, and ET indicate high segmentation accuracy.
Clinical Implications
The DPEA-Net model provides a practical tool for automated glioma subregion delineation, which can streamline neuro-oncological workflows. Its efficiency and accuracy may facilitate better treatment planning and monitoring in clinical settings.
Conclusion
DPEA-Net represents a significant advancement in automated glioma segmentation, combining accuracy with computational efficiency for clinical application.
