Rapid detection of novel coronavirus infection pneumonia
In early 2020, coronavirus disease (covid-19) spread worldwide, causing the world to face a crisis of survival and health. The automatic detection of lung infection through computed tomography ( C T ) images provides great potential for strengthening the traditional health care strategy to cope with covid-19. However, segmentation of infected areas from CT sections faces several challenges, including high variability in infection characteristics and low-intensity contrast of infection with normal tissue. In addition, it is unrealistic to collect a large amount of data in a short time, which hinders the training of the depth model. In order to meet these challenges, this paper proposes a new covid-19 lung infection segmentation depth network, which can automatically identify the infected areas from chest CT slices. In the network, a parallel partial decoder is used to aggregate high-level features and generate global graphs. Then, implicit inverse attention and explicit edge attention are used to model and enhance the boundary representation. In addition, in order to alleviate the shortage of labeled data, a semi supervised sequence configuration framework based on random selection propagation strategy is proposed. This framework only needs a small number of labeled images and mainly uses unlabeled data. Semi supervised framework can improve learning ability and achieve higher performance. Extensive experiments on covid semiseg and CT show that the proposed inf net is superior to most tip segmentation models and improves the latest performance.
1. The texture, size and position of the infection in CT sections vary greatly, which is challenging to detect. For example, the combined area of the whole piece is too small, which may easily lead to false negative of the whole piece of CT.
2. The variance between classes is small. For example, GGO boundaries are often low in contrast, fuzzy in appearance, and difficult to identify.
3. Due to the emergency of covid-19, it is difficult to collect enough marker data for the training of depth model in a short time.
1. By using the parallel partial decoder (PPD) to aggregate the features from the high level, the combined features obtain the context information and generate a global map as the initial guidance area for the subsequent steps. In order to further explore the boundary clues, we use a set of implicitly repeated reverse attention (RA) modules and explicit edge attention guidance to establish the relationship between regions and boundaries.
2. A semi supervised segmentation system based on the propagation of random selection is introduced to alleviate the short time of labeled data.
3. A covid-19 semi supervised infection segmentation (covid semiseg) dataset was constructed, including 100 labeled CT slices from the covid-19 CT segmentation dataset and 1600 unlabeled images from the covid-19 CT acquisition dataset
Edge map predicted with standard binary cross entropy constraint and edge map of ground truth value (GT):
Parallel partial decoder (PPD)
Reverse Focus module (RA)
A random sampling strategy is used to gradually expand the training data set using unlabeled data. False labels were generated for unlabeled CT images using the procedure described in the following figure. Then the model is trained using the obtained CT images with false labels.
The semi information network was extended to a multi class pulmonary infection detection framework to provide more abundant information for the further diagnosis and treatment of covid-19. The extension of the semi information network is based on the multi class marker framework guided by the infected area, as shown in the figure.