Regional-aware and sequence-informed multi-decoder network for robust brain glioma segmentation in multi-parametric...

Area of Science:

• Neuro-oncology
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Accurate glioblastoma subregion segmentation from multi-parametric MRI is crucial for patient care.
• Challenges exist in delineating tumor subregions due to heterogeneous imaging characteristics.
• Current deep learning models often underutilize the clinical specificity of individual MRI sequences.

Purpose of the Study:

• To develop a novel deep learning framework for accurate and robust glioblastoma subregion segmentation.
• To improve the delineation of enhancing tumor, non-enhancing tumor core, and peritumoral edema.
• To enhance the generalizability of segmentation models across diverse datasets.

Main Methods:

• A multi-decoder deep learning architecture for independent segmentation of key tumor subregions.
• A sequence-informed guidance strategy to align MRI sequences with specific diagnostic targets.
• A modified self-attention mechanism for improved feature recalibration and anatomical coherence.

Main Results:

• Achieved an average Dice Similarity Coefficient (DSC) of 0.9009 and HD95 of 6.61 mm on the BraTS 2023 dataset.
• Outperformed state-of-the-art methods, especially in enhancing tumor segmentation.
• Demonstrated strong generalizability across four external datasets, with DSC gains up to 4.09% in challenging scenarios.

Conclusions:

• The proposed framework offers a robust and generalizable solution for glioblastoma segmentation.
• Integration of clinical insight and methodological innovation enhances segmentation precision.
• Supports improved personalized treatment planning and outcome assessment in neuro-oncology.