Domain Generalized Medical Image Segmentation Based on Luminance Contrast Consistency

In the field of domain generalization medical image segmentation, most of the current data augmentation methods are considered from the shape of the object, and the data of different data distribution types can be obtained by changing the shape. However, the fact that medical images are predominantly gray and white has been ignored, which in turn has led to a lack of attempts to reduce the differences between the data by adjusting the luminance contrast of the images. In this paper, an adaptive luminance contrast consistency module (ALCC) is proposed to eliminate the luminance contrast differences of the data in each domain and improve the generalization ability of the model. Multi-scale feature learning is widely used for network modeling. To a large extent, it has promoted the development of medical image segmentation, however, these attention methods can weaken the network's perception of small objects, discard the features of small objects, and reduce the segmentation performance of the model. In addition, existing multiscale methods simply focus on views of different sizes, and their sparse multiscale feature collection is not rich enough for small target segmentation. In this paper, we propose a cascaded multiscale convolutional multidimensional attention module (CMCMA), which is incorporated into each skip-connection path to capture multiscale features with different semantic depths for the model's small-target-awareness capability. Our method exhibits significant performance on the ATLAS2.0 public dataset.