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

Source: arXiv - 2603.09945v1

Overview

The paper introduces k‑MTR, a novel end‑to‑end framework that learns directly from undersampled MRI k‑space data, bypassing the traditional “reconstruct‑then‑analyze” pipeline. By aligning raw frequency measurements with fully‑sampled image representations in a shared latent space, the method extracts diagnostic cardiac information (e.g., phenotypes, disease labels, segmentations) without ever forming a high‑resolution image first.

Key Contributions

• Task‑aware k‑space encoder that maps undersampled frequency data into a semantic latent space aligned with fully‑sampled images.
• Multi‑task learning across regression, classification, and segmentation, demonstrating that a single latent representation can serve diverse downstream cardiac analyses.
• Large‑scale simulation study (≈42 k virtual subjects) that validates the approach under realistic undersampling patterns and noise levels.
• Competitive performance against state‑of‑the‑art image‑domain baselines, showing no loss—and in some cases gains—in diagnostic accuracy despite operating on incomplete k‑space.
• Architectural blueprint for integrating k‑space representation learning into clinical cardiac MRI workflows, potentially reducing acquisition time and computational overhead.

Methodology

1. Data Simulation – The authors generated a synthetic cohort of 42 000 cardiac MRI subjects, each with fully sampled k‑space and corresponding ground‑truth images, phenotypic measurements, disease labels, and segmentation masks.
2. Dual‑branch Encoder –
   • k‑space branch: a convolutional network processes the undersampled frequency data (complex‑valued) and outputs a latent vector.
   • Image branch: a parallel encoder ingests the fully‑sampled image and produces a latent vector in the same dimensionality.
3. Latent Alignment Loss – A contrastive/metric loss forces the two latent vectors (k‑space vs. image) for the same subject to be close, while pushing apart vectors from different subjects. This aligns the low‑dimensional physiological semantics across modalities.
4. Multi‑Task Heads – The shared latent space feeds three downstream heads:
   • Regression (continuous cardiac phenotypes)
   • Classification (binary/multi‑class disease detection)
   • Segmentation (pixel‑wise anatomical masks)
5. Training Strategy – End‑to‑end optimization jointly minimizes the alignment loss and the task‑specific losses, allowing the k‑space encoder to learn representations that are already “task‑ready.”

The whole pipeline runs on standard deep‑learning frameworks (PyTorch/TensorFlow) and can be trained on commodity GPUs.

Results & Findings

• Accuracy parity: Across all three tasks, k‑MTR matches or slightly exceeds the performance of pipelines that first reconstruct a full image and then analyze it.
• Robustness to undersampling: Even with 4× acceleration (i.e., only 25 % of k‑space samples), the latent space retains enough anatomical detail for reliable downstream predictions.
• Speed & memory: Skipping the inverse Fourier reconstruction reduces per‑scan processing time by ~30 % and cuts GPU memory usage because the network never needs to handle full‑resolution images.

These findings validate the core hypothesis: the diagnostic information lives in a low‑dimensional manifold that can be accessed directly from raw k‑space.

Practical Implications

• Faster MRI protocols – Clinicians could acquire heavily undersampled scans, shortening exam time and improving patient comfort without sacrificing diagnostic quality.
• Edge‑device deployment – Since the model works on compact latent vectors, it can be integrated into scanner consoles or cloud‑edge pipelines, enabling real‑time decision support.
• Reduced reconstruction artifacts – By avoiding the ill‑posed inverse problem, common artifacts (e.g., Gibbs ringing, aliasing) that can mislead downstream AI tools are eliminated.
• Unified workflow – One model serves multiple downstream tasks (risk scoring, segmentation for planning, phenotype extraction), simplifying software stacks and maintenance for imaging vendors.
• Data efficiency – The latent alignment approach could be extended to other modalities (e.g., brain MRI, CT sinograms) where raw acquisition data is available but reconstruction is costly.

Limitations & Future Work

• Synthetic training data – The large‑scale cohort is simulated; real‑world validation on diverse scanner hardware and patient populations is still needed.
• Complex‑valued handling – Current implementation treats real and imaginary parts as separate channels; more sophisticated complex‑valued networks might improve performance.
• Generalization to other anatomies – While the cardiac domain is a natural testbed, extending k‑MTR to multi‑slice or 3D volumetric acquisitions will require architectural scaling.
• Interpretability – The latent space is highly abstract; future work could explore disentanglement techniques to make the learned representations more clinically interpretable.

Authors

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

Paper Information

• arXiv ID: 2603.09945v1
• Categories: cs.CV, cs.AI
• Published: March 10, 2026
• PDF: Download PDF