Reconstructing detail differently
Standard segmentation decoders rebuild spatial detail through conventional upsampling operations. This research asks whether that reconstruction can instead be learned and organised in the frequency domain.
I designed and implemented a PsychoNet-based decoder and derived Fourier Shuffle, a frequency–spatial upsampling operation for medical image segmentation. The architecture was implemented in PyTorch and evaluated with custom training scripts on GPU compute nodes.
Results
Against a state-of-the-art U-Net baseline on the Synapse dataset, the final model achieved a 1.77% improvement in validation performance while using 13.8% less GPU memory.
The project began during my UQ Summer Research Scholarship and led to a separate Research Assistant role. I am now extending the mathematical ideas through a Generalised Kaleidoscope transform for frequency-based autoencoders.
Collaboration
This work was supervised by Shekhar Chandra and Wendi Ma, with research collaboration from Youssef Hassan. The broader work contributes to ongoing research in frequency-domain methods for semantic vision.