Abstract

This paper explores the embedding of a Fourier-Feature—enhanced multiplayer perceptron(MLP-FEE) at the heart of a newly refactored python workflow for four-dimensional cardiac-MRI strain quantification demonstrating how a single, compact network can outperform traditional convolution and spline-based methods. The original code, capable of orientation normalization, displacement tracking, and finite-difference strain computation, has been translated and consolidated into pytorch. By injecting sinusoidal positional encodings at the network’s input layer supplied a rich set of high-frequency basis functions hence enabling multilayer MLP to resolve gradients that cubic splines and conventional CNNs typically blur or struggle with. Profiling on an Apple-silicon GPU shows interactive inference without hand-tuned CUDA alongside a network has a higher precision while trimming code complexity by roughly 50%. Qualitative comparison with finite-element ground truths confirms preservation of peak systolic strain and segment-wise GLS trends. Meanwhile, the transpilation preserves domain-specific algorithms while exposing them to modern deep-learning tooling along with swappable optimisers and augmentation strategies. By showcasing the power of Fourier features inside a lightweight MLP, this work illustrates a path for researchers to migrate legacy workflows into a GPU-accelerated, open-source ecosystem to unlock higher-resolution insights across a variety of Biomedical Engineering datasets.

Keywords

Fourier Feature MLP, JAX, GPU Acceleration, Biomedical Imaging, Machine Learning

Date of this Version

8-5-2025

Share

COinS