Improving u-net performance for low-grade glioma MRI image segmentation by modification of u-net concatenations
Ziheng Su, Tario You · 2024
Recent developments in deep learning techniques applied to lower-grade glioma (LGG) MRI image segmentation have created accurate models comparable to human professionals. Among these deep learning models, the u-net proposed by Buda et al. performs best among other model architectures. However, we doubt that u-net reaches its full potential. In this paper, we validate whether u-net has the maximum performance possible: we modify the number and direction of concatenation operations in u-net to see whether modified architectures perform better than the original u-net in terms of accuracy, measured by dice coefficients. Our results show 1) the original u-net model is not optimal since all modified models perform better than the original one, 2) positive and likely logarithmic correlation between concatenation complexity and model performance, and 3) unknown relation between concatenation direction and model performance. The reason for the logarithmic part of the correlation in 2) and the unknown correlation in 3) requires further investigation with more theories.