Segmenting Brain Tumor Using Cascaded V-Nets in Multimodal MR Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate segmentation of brain tumor substructures is crucial for diagnosis and treatment planning.
• Existing V-Net models show promise but can be improved for complex segmentation tasks.

Purpose of the Study:

• To develop and evaluate a novel cascaded V-Nets method for enhanced brain tumor segmentation.
• To improve segmentation accuracy by employing ensemble strategies and specialized preprocessing pipelines.

Main Methods:

• A cascaded V-Net architecture with skip connections and focal loss was implemented.
• Three distinct preprocessing pipelines were used to train multiple models.
• An ensemble strategy combined segmentation probability maps from different models.
• A hierarchical segmentation approach was adopted: whole tumor first, then substructures.

Main Results:

• Achieved high Dice scores on the BraTS 2018 dataset: 0.9048 (whole tumor), 0.8364 (tumor core), and 0.7748 (enhancing tumor).
• Validated performance on local hospital datasets, yielding comparable results.
• Demonstrated potential for patient overall survival prediction using ensemble classifiers.

Conclusions:

• The proposed cascaded V-Nets method with ensemble strategy significantly improves brain tumor substructure segmentation.
• The hierarchical segmentation approach is effective for delineating tumor components.
• The method shows promise for both segmentation and survival prediction in neuro-oncology.