Progressive attention module for segmentation of volumetric medical images.

Purpose: Medical image segmentation is critical for many medical image analysis applications. 3D convolutional neural networks (CNNs) have been widely adopted in the segmentation of volumetric medical images. The recent development of channelwise and spatialwise attentions achieves the state-of-the-art feature representation performance. However, these attention strategies have not explicitly modeled interdependencies among slices in 3D medical volumes. In this work, we propose a novel attention module called progressive attention module (PAM) to explicitly model the slicewise importance for 3D medical image analysis.

Methods: The proposed method is composed of three parts: Slice Attention (SA) block, Key-Slice-Selection (KSS) block, and Channel Attention (CA) block. First, the SA is a novel attention block to explore the correlation among slices for 3D medical image segmentation. SA is designed to explicitly reweight the importance of each slice in the 3D medical image scan. Second, the KSS block, cooperating with the SA block, is designed to adaptively emphasize the critical slice features while suppressing the irrelevant slice features, which helps the model focus on the slices with rich structural and contextual information. Finally, the CA block receives the output of KSS as input for further feature recalibration. Our proposed PAM organically combines SA, KSS, and CA, progressively highlights the key slices containing rich information for the relevant tasks while suppressing those irrelevant slices.

Results: To demonstrate the effectiveness of PAM, we embed it into 3D CNNs architectures and evaluate the segmentation performance on three public challenging data sets: BraTS 2018 data set, MALC data set, and HVSMR data set. We achieve 80.34%, 88.98%, and 84.43% of the Dice similarity coefficient on these three data sets, respectively. Experimental results show that the proposed PAM not only boosts the segmentation accuracy of the standard 3D CNNs methods consistently, but also outperforms the other attention mechanisms with slight extra costs.

Conclusions: We propose a new PAM to identify the most informative slices and recalibrate channelwise feature responses for volumetric medical image segmentation. The proposed method is evaluated on three public data sets, and the results show improvements over other methods. This proposed technique can effectively assist physicians in many medical image analysis. It is also anticipated to be generalizable and transferable to assist physicians in a wider range of medical imaging applications to produce greater value and impact to health.