From Anatomy to Genomics Using a Multi-Task Deep Learning Approach for Comprehensive Glioma Profiling.

Area of Science:

• Neuro-oncology
• Medical Imaging
• Genomics
• Artificial Intelligence

Background:

• Gliomas are complex and lethal primary brain tumors.
• Accurate assessment of anatomical subregions and molecular alterations is crucial for glioma management.
• Current bioimage analysis pipelines often treat MRI segmentation and molecular biomarker prediction as separate tasks, leading to inefficiencies.

Purpose of the Study:

• To develop an integrated deep learning framework for simultaneous anatomical segmentation and molecular biomarker prediction in gliomas.
• To overcome the disconnected nature of existing bioimage analysis pipelines.
• To enhance precision and reduce diagnostic latency in neuro-oncology.

Main Methods:

• Introduced MGMT-Net, a novel deep learning scheme utilizing multi-modal MRI data without conversion.
• Incorporated a Cross-Modality Attention Fusion (CMAF) module for integrating diverse imaging sequences.
• Employed a hybrid Transformer-CNN encoder for capturing global and local details, with dual-task decoders for segmentation and genomic classification.

Main Results:

• MGMT-Net demonstrated high segmentation accuracy and robust biomarker classification performance on BraTS 2024 and TCGA/EGD datasets.
• The model showed strong generalizability across external institutional cohorts.
• Ablation studies validated the contribution of each architectural component to the model's robustness.

Conclusions:

• MGMT-Net offers a scalable, clinically relevant solution integrating radiological imaging and genomic insights.
• The AI-driven approach has the potential to reduce diagnostic latency and improve precision in neuro-oncology.
• This work advances comprehensive, AI-driven glioma assessment by integrating spatial and genetic analysis.