DLTTA

DLTTA implements dynamic learning-rate modulation for test-time adaptation to mitigate cross-domain distribution shifts in medical image analysis.

Key Features:

• Dynamic Learning Rate Adjustment: Modulates the learning rate for each test image individually to accommodate distribution shifts in sequential test data during Test-time Adaptation (TTA).
• Memory Bank-Based Estimation: Implements a memory bank-based estimation scheme that quantifies the distribution discrepancy of each test sample relative to the training data.
• Adaptive Weight Updates: Dynamically adjusts learning rates based on estimated discrepancies to determine the appropriate degree of weight updates for each test sample.

Scientific Applications:

• Retinal Optical Coherence Tomography (OCT) Segmentation: Applied to retinal OCT segmentation to adapt segmentation models to domain-specific variations.
• Histopathological Image Classification: Applied to histopathological image classification to adapt classifiers across data from different sources.
• Prostate 3D MRI Segmentation: Applied to prostate 3D MRI segmentation to manage cross-domain discrepancies in volumetric MRI data.

Methodology:

Per-image dynamic learning-rate modulation during test-time adaptation; memory bank-based estimation of sample-to-training distribution discrepancy; adaptive adjustment of learning rates to control weight updates.