Analyzing the Effect of $k$-Space Features in MRI Classification Models
Pascal Passigan, Vayd Ramkumar
TL;DR
The paper tackles opacity in MRI classification by integrating $k$-space ($k$-space) features with spatial-domain CNN inputs and using $UMAP$ to visualize latent embeddings, aiming to improve interpretability. It introduces a dual-domain CNN that takes a three-channel input comprising the spatial image and the real and imaginary parts of the Fourier-transformed data, enabling frequency-domain information to inform decision-making. Results indicate faster training and more separable latent clusters in $UMAP$, but also reveal increased confusion between closely related dementia categories, highlighting a trade-off between interpretability and discriminative precision. The work suggests that frequency-domain augmentation, paired with $UMAP$-based interpretability, has potential clinical value in radiology and motivates further refinement to balance accuracy and transparency.
Abstract
The integration of Artificial Intelligence (AI) in medical diagnostics is often hindered by model opacity, where high-accuracy systems function as "black boxes" without transparent reasoning. This limitation is critical in clinical settings, where trust and reliability are paramount. To address this, we have developed an explainable AI methodology tailored for medical imaging. By employing a Convolutional Neural Network (CNN) that analyzes MRI scans across both image and frequency domains, we introduce a novel approach that incorporates Uniform Manifold Approximation and Projection UMAP] for the visualization of latent input embeddings. This approach not only enhances early training efficiency but also deepens our understanding of how additional features impact the model predictions, thereby increasing interpretability and supporting more accurate and intuitive diagnostic inferences