Attention U-Net: Learning Where to Look for the Pancreas

The paper introduces an attention gate (AG) model for medical imaging that automatically learns to focus on target structures of varying shapes and sizes. Models trained with AGs implicitly learn to suppress irrelevant regions of an input image while highlighting salient features useful for a task, which eliminates the need for the explicit external tissue or organ localisation modules that cascaded CNNs typically require. AGs can be integrated into standard architectures such as U-Net with minimal computational overhead.

The resulting Attention U-Net increases model sensitivity and prediction accuracy, and it was evaluated on two large CT abdominal datasets for multi-class image segmentation. Experimental results show that attention gates consistently improve U-Net's prediction performance across different datasets and training set sizes while preserving computational efficiency, and the authors made the code publicly available.