An interpretable feature-learned model for overall survival classification of High-Grade Gliomas

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• High Grade Gliomas (HGGs) are aggressive brain tumors with high incidence.
• Accurate survival prediction is crucial for effective HGG patient management.
• Current prediction methods require improvement for better clinical outcomes.

Purpose of the Study:

• To develop a unified framework for fully automatic overall survival classification of High Grade Gliomas.
• To interpret the survival prediction for better clinical understanding.
• To improve the accuracy and reliability of HGG survival prognostication.

Main Methods:

• A glioma detection model identifies tumorous images.
• A pre-processing module extracts 2D slices and creates survival data arrays.
• A classification pipeline uses modality-specific and modality-concatenated pathways with CNNs to extract features from HGG sub-regions (edema, enhancing tumor, necrosis) across multiple neuro-imaging modalities.

Main Results:

• The proposed framework achieved outstanding classification performance on BraTS 2018 and BraTS 2019 datasets.
• Classification accuracy, sensitivity, and specificity reached approximately 0.998, 0.997, and 0.999 for BraTS 2018, and 1.000, 0.999, and 0.999 for BraTS 2019.
• The model successfully classified HGG patients into short, mid, and long survival groups.

Conclusions:

• The developed model demonstrates superior performance in overall survival classification for High Grade Gliomas.
• The unified framework provides highly accurate and interpretable survival predictions.
• This approach holds significant potential for improving HGG patient care and treatment planning.