Attention U-net: Learning Where to Look for the pancereas
Address: arxiv.org/abs/1804.03…
Github:github.com/ozan-oktay/…
AGS attention Gates for Image Analysis
Legacy: Still don’t know how to do it after reading it
Abstract: We present a self-attention Gating module for attention Gate for medical imaging, which automatically learns to distinguish the shape and size of the target. This attention-gate model is trained to suppress irrelevant areas and focus on useful salient features, which is effective for a specific task.
The cascade framework extracts ROI from the region of interest and makes intensive predictions about that particular ROI. However, this approach leads to excessive and redundant use of computational resources and model parameters. For example, all models in a cascade repeatedly extract similar low-level features.
To solve this problem, attention Gates (AGS) is proposed in this paper, which has the following advantages:
1. The grid-based AG is proposed to make the attention coefficient more prominent in local areas. CNN with AG allows you to learn end-to-end, without the need for additional monitoring standards.
2. Automatically focus on areas with distinctive features in reasoning
3, will not introduce a lot of parameters and calculation. AG improves model sensitivity and accuracy by inhibiting activation of features in unrelated regions for intensive label prediction. Thus, the need to use external organ localization models can be eliminated while maintaining high accuracy. In this paper, the segmentation method is divided into two steps, namely detection and segmentation. ROI of the organs and tissues to be segmented is determined first, and then small regions are segmented.
4. Soft attention was used in CNN of medical images for the first time. This module can replace hard attention in classification task and positioning module in organ positioning task.
5. U-net was improved to attention U-NET, which increased the sensitivity of the model to foreground pixels, and the design experiment proved that this improvement was universal.
Related work:
CT pancreas segmentation: There are mainly statistical Shape Models, multi-Atlas techniques and cascaded multi-state CNN models. In CNN, an initial coarse-grained model (such as U-Net or regression forest) is used to obtain ROI, then the cropped ROI is used for subdivision through a second model, and a combination of 2D-FCN and recursive neural network (RNN) model is used to exploit the dependence between adjacent axial slices.
The specific methods
FCN(fully convolutional network) fully convolutional network