End-to-End Multi-task Learning Architecture for Brain Tumor Analysis with Uncertainty Estimation in MRI Images

Area of Science:

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

Background:

• Gliomas are aggressive brain tumors with complex structures, leading to diagnostic challenges and errors.
• Manual segmentation of gliomas is time-consuming and prone to subjectivity.
• Existing AI solutions lack an end-to-end system for comprehensive glioma analysis.

Purpose of the Study:

• To develop an integrated AI framework for automated glioma classification, segmentation, and survival prediction.
• To leverage multi-task learning (MTL) and feature attention to improve diagnostic accuracy and efficiency.
• To incorporate uncertainty estimation to enhance clinician confidence in AI-driven results.

Main Methods:

• An end-to-end multi-task learning (MTL) architecture with a feature attention module was developed.
• The framework integrates classification, segmentation, and survival prediction tasks.
• Experiments utilized Brain Tumor Segmentation (BraTS) 2019 and 2020 datasets with various MRI sequences.

Main Results:

• The best model achieved 95.1% accuracy for classification and an 86.3% dice score for segmentation.
• Survival prediction yielded a Mean Absolute Error (MAE) of 456.59.
• Uncertainty estimation demonstrated that increased data improves generalization and accuracy.

Conclusions:

• Deep learning-based MTL models can automate brain tumor analysis, offering efficient and accurate results with minimal inference time.
• The proposed framework shows potential for clinical application in initial glioma patient screening.
• Uncertainty quantification is crucial for improving model reliability and generalization ability.