No Image, No Problem: End-to-End Multi-Task Cardiac Analysis Straight from Undersampled k-Space

Yundi Zhang, Sevgi Gokce Kafali, Niklas Bubeck, Daniel Rueckert, Jiazhen Pan

Keywords: Cardiac MRI, multi-task, k-space measurements, representation learning, foundation models

TL;DR: We propose k-MTR, the first framework to enable unified multi-task cardiac analysis directly from undersampled k-space, bypassing image reconstruction entirely.

Abstract: Before a cardiac MR image is reconstructed, the heart is represented in k-space, which encodes all information needed for analysis—including tissue structure, motion, and functional dynamics. However, information extraction (e.g., downstream analysis tasks such as segmentation or biomarker quantification) is usually performed in the image domain rather than in the k-space domain. This means that the quality of the information extraction is fundamentally limited by the image reconstruction quality. At the same time, the push toward unified models for diverse cardiac downstream tasks has accelerated, driven by advances in efficient representation learning. However, most work has focused on the image domain, overlooking the k-space potential as a direct, information-dense source for end-to-end, multi-task cardiac analysis. As a result, the development of robust and expressive k-space representations and their impact on downstream cardiac assessment remain significantly underexplored. To address this gap, we propose $\textbf{k}$-space $\textbf{M}$ulti-$\textbf{T}$ask $\textbf{R}$epresentation ($\textbf{k-MTR}$) learning, which enables solving different downstream tasks directly from undersampled k-space. By aligning the k-space and image-domain embeddings, k-MTR establishes a unified representation that simultaneously captures local anatomical detail, global spectral structure, and rich physiological signatures. We show that k-MTR matches or exceeds state-of-the-art image-based and k-space–based baselines across three clinically relevant tasks-disease classification, phenotype regression, and segmentation, providing the first systematic evidence that k-space alone can support comprehensive cardiac analysis. k-MTR represents a pivotal step toward scalable, reconstruction-free cardiac foundation models. The code will be made publicly available after the review process.

Primary Subject Area: Unsupervised Learning and Representation Learning

Secondary Subject Area: Application: Cardiology

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 143