Attention Augmented U-Net for Robust Thyroid Ultrasound Segmentation

Abstract

Thyroid nodule segmentation in ultrasound imaging faces challenges from speckle noise, low contrast, and morphological variability. We propose an EfficientNet-B5-backboned UNet++ architecture enhanced with Dual-Path Attention Modules (DPAM) integrating Global Context, Efficient Channel, and Spatial Attention mechanisms. Our model employs a composite loss function combining Focal Tversky, Dice, and Boundary losses with temporal ramp-up strategy to address class imbalance. Training features a three-phase progressive schedule with automatic threshold optimization and progressive resizing to 512×512, complemented by ultrasound-specific data augmentations. Evaluated on the TN3K dataset using 5-fold cross-validation, our model achieves state-of-the-art performance: 0.8875 Dice, 0.7969 IoU, 0.9688 accuracy, representing improvement over recent CNN and Transformer baselines while maintaining computational efficiency with less than 50M parameters. The tight training-validation gaps demonstrate robust generalization, indicating reliable boundary delineation and clinical applicability for computer-aided diagnosis pipelines.

Ähnliche Arbeiten