MT-IDS: A multi-task information decoupling strategy for identifying lymph node metastasis in the mediastinal region

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate identification of metastatic lymph nodes in mediastinal regions is crucial for lung cancer staging.
• Traditional single-task algorithms and existing multi-task learning methods face challenges in handling multi-dimensional classification tasks and balancing feature relationships.

Purpose of the Study:

• To propose a novel Multi-Task Information Decoupling Strategy (MT-IDS) to address the limitations of existing methods in multi-dimensional medical image classification for lung cancer staging.

Main Methods:

• MT-IDS decomposes the main task into auxiliary tasks across different feature dimensions for unified optimization.
• A Dual-control Branch Routing Gate Mechanism (DBR) is used for precise weighting of shared and task-specific features.
• A Dual-Dimensional Gradient Balancing Algorithm (DD-GB) ensures gradient alignment and dynamic scaling for inter-task balance.

Main Results:

• MT-IDS demonstrated significant advantages in both ablation and comparative experiments.
• The proposed strategy effectively optimizes detection performance across multiple tasks in lung cancer staging.

Conclusions:

• MT-IDS offers an innovative solution for multi-dimensional medical image classification problems.
• The strategy shows potential for improving the accuracy and efficiency of lung cancer staging through enhanced lymph node metastasis assessment.