[Paper Review] AUNet: Attention-guided dense-upsampling networks for breast mass segmentation in whole mammograms

Hui Sun, Cheng Li+4 more|arXiv (Cornell University)|Oct 24, 2018

AI in cancer detection76 references18 citations

TL;DR

This paper proposes AUNet, a novel attention-guided dense-upsampling network for direct breast mass segmentation in whole mammograms, avoiding manual or automatic patch extraction. By integrating an asymmetrical encoder-decoder with an attention-guided dense-upsampling block (AU block) that enhances feature fusion and preserves spatial details, AUNet achieves state-of-the-art performance with Dice scores of 81.8% on CBIS-DDSM and 79.1% on INbreast.

ABSTRACT

Mammography is one of the most commonly applied tools for early breast cancer screening. Automatic segmentation of breast masses in mammograms is essential but challenging due to the low signal-to-noise ratio and the wide variety of mass shapes and sizes. Existing methods deal with these challenges mainly by extracting mass-centered image patches manually or automatically. However, manual patch extraction is time-consuming and automatic patch extraction brings errors that could not be compensated in the following segmentation step. In this study, we propose a novel attention-guided dense-upsampling network (AUNet) for accurate breast mass segmentation in whole mammograms directly. In AUNet, we employ an asymmetrical encoder-decoder structure and propose an effective upsampling block, attention-guided dense-upsampling block (AU block). Especially, the AU block is designed to have three merits. Firstly, it compensates the information loss of bilinear upsampling by dense upsampling. Secondly, it designs a more effective method to fuse high- and low-level features. Thirdly, it includes a channel-attention function to highlight rich-information channels. We evaluated the proposed method on two publicly available datasets, CBIS-DDSM and INbreast. Compared to three state-of-the-art fully convolutional networks, AUNet achieved the best performances with an average Dice similarity coefficient of 81.8% for CBIS-DDSM and 79.1% for INbreast.

Motivation & Objective

• To address the challenge of accurate breast mass segmentation in whole mammograms with low signal-to-noise ratios and variable mass shapes and sizes.
• To eliminate the need for manual or automatic image patch extraction, which introduces errors and is time-consuming.
• To improve feature representation and spatial detail recovery in the segmentation process through enhanced upsampling and attention mechanisms.
• To develop a fully convolutional network that directly processes whole mammograms without patch-based preprocessing.
• To achieve superior segmentation performance on benchmark datasets using a novel upsampling block with attention and dense connections.

Proposed method

• AUNet employs an asymmetrical encoder-decoder architecture to effectively capture multi-scale features from whole mammograms.
• The core innovation is the attention-guided dense-upsampling block (AU block), which combines dense upsampling with channel attention to reduce information loss.
• The AU block fuses high-level semantic features with low-level spatial details more effectively than standard upsampling methods.
• Channel attention is integrated to emphasize informative feature channels, improving representation quality.
• Dense skip connections within the AU block enhance feature propagation and gradient flow during training.
• The network is end-to-end trained using binary cross-entropy and dice loss to optimize segmentation performance.

Experimental results

Research questions

• RQ1 Can a fully convolutional network achieve superior breast mass segmentation in whole mammograms without relying on patch extraction?
• RQ2 How does attention-guided dense-upsampling improve feature recovery and segmentation accuracy compared to standard upsampling?
• RQ3 To what extent does the integration of channel attention enhance the discriminative power of feature maps in mass segmentation?
• RQ4 How does AUNet compare to state-of-the-art fully convolutional networks on public mammogram datasets?
• RQ5 Can the proposed AU block generalize across diverse mass shapes, sizes, and noise levels in clinical mammograms?

Key findings

• AUNet achieved a mean Dice similarity coefficient of 81.8% on the CBIS-DDSM dataset, outperforming three state-of-the-art fully convolutional networks.
• On the INbreast dataset, AUNet achieved a Dice score of 79.1%, demonstrating strong generalization across different imaging protocols and masses.
• The ablation study confirmed that each component of the AU block—dense upsampling, attention mechanism, and feature fusion—contributed significantly to performance gains.
• The network demonstrated robustness to low-contrast and noisy regions typical in mammograms, preserving fine boundaries of masses.
• The attention mechanism successfully highlighted informative channels, improving segmentation accuracy in complex anatomical regions.
• The direct whole-mammogram inference capability of AUNet eliminated the error propagation associated with patch-based approaches.