The paper I read this week is a paper titled
"C2FNAS: Neural Structure Search from Coarse to Fine for 3D Medical Image Segmentation" published on CVPR in 2020 .
Professor Alan L. Yuille is an outstanding professor of cognitive science and computer science at Johns Hopkins University and one of the founders of computer vision. He leads research groups on component cognition, vision and learning. He moved to Johns Hopkins University in January 2016. His research interests include visual computing models, cognitive mathematical models, medical image analysis, and artificial intelligence and neural networks.
The above right image Figure 1 shows some examples of images and masks related to tumors in the MSD medical image segmentation task. Image anomalies, texture changes, and anisotropy will cause great challenges. Red, green, and blue correspond to labels 1, 2, and 3 of each data set, respectively.
The author is analyzing the memory consumption problem caused by the large-scale network of the 3D medical image segmentation task. It is not feasible to sacrifice the input size alone, which usually leads to the training problem of unstable convergence. In order to solve this problem, combining NAS and medical image segmentation, a neural structure search algorithm from coarse to fine is proposed to automatically design a three-dimensional segmentation network.
Figure 3 shows an example of how to help reduce the search space by introducing a priori. The gray nodes will be completely eliminated from the graph. In addition, many illegal paths have been cleared. Examples of illegal paths and legal paths are shown as orange line paths and green line paths, respectively. This reduces the search space to make the search process more focused and efficient.
It can be seen in Table 1 that the MSD challenge test set was measured by DSC coefficients and compared with the most advanced methods. *Indicates 5-layer model integration. The author also reports the average values of tasks and goals separately, so as to make an overall comparison of all tasks/objectives. From the average value, it can be seen that the overall performance and robustness are still very good.
The author's model shows better performance than the most advanced methods on most tasks, especially challenging tasks, and has a smaller model size compared to the most popular 3D models (see Table 2) . It is worth noting that the previous nnU-Net used various data expansion and test time expansion to improve performance, while the author only used simple rotation and flipping data expansion instead of test time expansion. Small data sets (such as the heart and hippocampus data sets) rely more on enhancement, while powerful architectures are prone to overfitting, which explains why the author's model does not outperform competitors on these data sets. In addition, nnU-Net uses different networks and hyperparameters for each task, while the author uses the same model and hyperparameters for all tasks, which shows that the author's model is not only more powerful, but also more robust and more generalized.
Left Table 2 compares FLOP parameters and can be seen with other 3D network, the author of more lightweight model.
Figure 6 provides some visual comparisons. Visual comparison of the latest methods (groups 1 and 2) with C2FNAS-PANC on the MSD test set. Visualize a case from the three most challenging tasks: pancreatic and pancreatic tumors, colon cancer and lung cancer. Red indicates abnormal pancreas, colon cancer and lung cancer, and green indicates pancreatic tumors. The case ID and dice score of C2FNAS-PANC are at the bottom, which is very effective.
Table 4 below shows the effect of model scaling. The numbers in the first column indicate the scale factors applied to the C2FNAS-PANC model. The results are based on the single-layer validation set and the final pancreatic data set search model.
