Memory-Augmented Capsule Network for Adaptable Lung Nodule Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Computer-aided diagnosis (CAD) systems face challenges adapting to new data distributions from diverse sources.
• Re-training CAD models for new domains typically demands extensive labeled data, which is costly and time-consuming.

Purpose of the Study:

• To develop a memory-augmented capsule network for efficient and rapid adaptation of CAD models to unseen domains.
• To reduce the reliance on large labeled datasets for CAD system re-training.

Main Methods:

• A capsule network extracts feature embeddings from high-dimensional input.
• A memory-augmented task network leverages stored knowledge from target domains.
• The proposed network adapts to new domains using minimal annotated samples.

Main Results:

• Achieved clinically relevant performance (AUROC 0.925 and 0.891) with only 30 additional samples on new datasets.
• Demonstrated efficiency equivalent to using two orders of magnitude less labeled data.
• Maintained robust performance (AUROC > 0.7) against noise, artifacts, and adversarial attacks, outperforming state-of-the-art methods.

Conclusions:

• The memory-augmented capsule network enables rapid and data-efficient adaptation of CAD systems.
• The method shows significant promise for improving the practical deployment and robustness of CAD systems in diverse clinical settings.