Internal-transfer Weighting of Multi-task Learning for Lung Cancer Detection

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Deep learning excels in medical imaging tasks like lung nodule segmentation and cancer prediction using computed tomography (CT).
• Multi-task learning networks can enhance performance on primary tasks by leveraging auxiliary tasks, but balancing optimization criteria remains challenging.

Purpose of the Study:

• To extend a 3D attention-based network with multi-task subnetworks for lung cancer detection and auxiliary respiratory disease diagnoses.
• To introduce and evaluate a Periodic Focusing Learning Policy (PFLP) and an Internal-Transfer Weighting (ITW) strategy for optimizing multi-task learning.

Main Methods:

• A 3D attention-based network was augmented with four auxiliary tasks: asthma, chronic bronchitis, COPD, and emphysema diagnosis.
• The Periodic Focusing Learning Policy (PFLP) was implemented to alternate task dominance during training.
• The Internal-Transfer Weighting (ITW) strategy was proposed to suppress auxiliary task losses in later training stages.
• Experiments utilized data from the National Lung Screening Trial (NLST) and Vanderbilt Lung Screening Program (3386 patients total).

Main Results:

• The baseline single-task network achieved an AUC of 0.8080 for lung cancer prediction.
• The multi-task baseline without adaptive weights failed to converge (AUC 0.6720).
• PFLP improved the multi-task network's lung cancer prediction AUC to 0.8402.
• The combination of PFLP and ITW further boosted performance to an AUC of 0.8462 (p < 0.01).

Conclusions:

• Adaptive weighting of multi-task learning is crucial for improving deep learning model performance in medical imaging.
• PFLP and ITW represent promising strategies for enhancing multi-task learning, particularly for primary tasks like lung cancer prediction.