Radon-AFMT Invariant Guided Conditional SE-UNet for MRI Tumor Segmentation

This paper proposes a brain tumor segmentation method that integrates analyti cally exact rotation, scale, and translation-invariant (RST-invariant) Radon-AFMT descrip tors into a U-Net architecture through Conditional Squeeze-and-Excitation (CSE) blocks. Unlike conventional attention mechanisms based solely on data-driven statistics, the pro posed approach conditions channel recalibration on precomputed geometric invariants, en abling geometry-aware feature modulation across encoder stages and the bottleneck. The architecture also includes residual convolutional blocks, attention gates on skip connections, and deep supervision for stable multi-scale training. Experiments on the Figshare Brain Tu mor dataset (3,064 MRI images) achieve a Dice score of 0.870, demonstrating the benefit of combining invariant signal representations with deep learning for robust medical image segmentation.