A radiomics-incorporated deep ensemble learning model for multi-parametric MRI-based glioma segmentation

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Accurate glioma segmentation on multi-parametric magnetic resonance imaging (mp-MRI) is crucial for diagnosis and treatment planning.
• Current segmentation methods may not fully capture the complex heterogeneity of gliomas.

Purpose of the Study:

• To develop and evaluate a deep ensemble learning (DEL) model incorporating radiomics spatial encoding for enhanced glioma segmentation accuracy using mp-MRI.
• To improve segmentation of enhancing tumor (ET), tumor core (TC), and whole tumor (WT).

Main Methods:

• A deep ensemble learning (DEL) model was developed using 369 glioma patients with T1, T1-Ce, T2, and FLAIR mp-MRI sequences.
• Radiomic features were extracted using a 3D sliding kernel, encoded as radiomic feature maps (RFMs), and reduced using principal component analysis (PCA).
• Four U-Net sub-models processed mp-MRI and PCA components, with results combined for segmentation of ET, TC, and WT.

Main Results:

• The radiomics-enhanced DEL models showed improved segmentation performance compared to mp-MRI-only U-Net.
• Dice coefficients increased for ET (0.777 to 0.817), TC (0.742 to 0.757), and WT (0.823 to 0.854).
• Accuracy, sensitivity, and specificity also demonstrated improvements.

Conclusions:

• Radiomics spatial encoding effectively captures image heterogeneity, enhancing segmentation accuracy.
• The proposed DEL model offers a novel and effective tool for mp-MRI-based glioma segmentation.
• This approach holds promise for improving clinical decision-making in neuro-oncology.