Learning Multi-Scale Attention Model for Spine Multi-Category Segmentation

Abstract

In the second CSIG spine multi-category segmentation challenge, the official spine structure data of 172 cases were provided, which contains up to 20 categories. If the multiscale method of encoder-decoder structure is used in multi-category segmentation at this dataset level, similar low-level features will be extracted multiple times, resulting in redundant use of information and the scale of the dataset limits the learning effect of the model. In order to avoid the aforementioned limitations, this work leverages a method based on a multiscale attention mechanism to solve the problem of multi-category segmentation. First, perform both operations of standardization and data augmentation on the given competition data, aiming to reinforce the data quality and the scalability. Secondly, the features at different scales are exploited through the Resnet network architecture, in parallel, the attention module consisting of the channel attention mechanism and the position attention mechanism extracts the features at different scales to obtain the corresponding attention maps. Finally, the features at different scales and the corresponding attention maps are fused in a weighted average to obtain the final prediction results. In the data set provided by the CSIG, Our multi-category segmentation method performance is 0.8438, which ranks 10-th place in the competition.