In many scientific domains, from medical imaging to fluid dynamics, we face a fundamental "inverse problem": we want to recover a high-resolution, complete signal from sparse, noisy, or undersampled measurements. Traditionally, deep learning solves this using supervised learning—training a model to map "bad" data to "perfect" ground-truth references.
However, as noted by Burhaneddin Yaman and Mehmet Akcakaya, acquiring fully-sampled ground truth is often physically impossible, such as when imaging a moving heart or capturing high-speed fluid turbulence. This has led to a powerful new paradigm: Measurement Splitting. This technique allows models to learn the underlying physics of a signal using only the incomplete data itself.
The Trivial Mapping Trap
If you try to train a neural network to reconstruct a signal using only undersampled measurements as both the input and the target, the network will quickly discover a "cheat code." It will learn the identity mapping—simply passing the input measurements to the output without filling in any missing information. To the network, the simplest way to minimize error is to change nothing.
To force the network to actually learn the structure of the data, we must hide some of the information.
The Mechanism of Disjoint Partitioning
The core intuition behind measurement splitting, formalized in Self-Supervised Learning of Physics-Based Reconstruction Neural Networks without Fully-Sampled Reference Data, is to divide the available undersampled measurements ($\Omega$) into two disjoint subsets:
- The Input Set ($\Theta$): A portion of the measurements used to generate a reconstruction.
- The Target Set ($\Lambda$): The remaining measurements, held out during the reconstruction process.
The loss function is then calculated only on the held-out set $\Lambda$. Because the network never sees the measurements in $\Lambda$ during the forward pass, it cannot simply copy them. To minimize the loss, it must learn to interpolate the missing values by understanding the global patterns and physical constraints of the signal. As Victor Sechaud and Julián Tachella argue in their work on Equivariant Splitting, this works because the underlying signal possesses redundancies (like sparsity or symmetry) that the measurement process does not.
Physics as the Interpolator
Measurement splitting doesn't work in a vacuum; it requires a model that "understands" the data's domain. Researchers like Kerstin Hammernik and Florian Knoll emphasize physics-driven deep learning, where the neural network architecture "unrolls" a traditional optimization algorithm.
By embedding the physical forward model (e.g., the Fourier transform in MRI) directly into the network, the model is constrained to produce images that are mathematically consistent with the sampled data. This "data consistency" layer ensures that the network's guesses for the missing $\Lambda$ points aren't just visually plausible, but physically accurate.
Dual-Domain Consistency
Recent advancements have moved beyond simple spatial interpolation. Bo Zhou and Neel Dey introduced DSFormer, which enforces consistency across both the image domain (spatial features) and the frequency domain (raw sensor data).
This dual-domain approach prevents the artifacts that often plague self-supervised models. If the model's reconstruction looks good in the image domain but contradicts the raw measurements in the frequency domain, the loss function penalizes it. This "cross-check" allows self-supervised models to reach performance levels nearly identical to supervised models trained on expensive ground truth.
Generalizing Across Modalities
While MRI has been the primary laboratory for these ideas, the principle of measurement splitting is remarkably general:
- Fluid Dynamics: Zihan Zhou demonstrated in Incomplete Data, Complete Dynamics that one can learn full physical fields (like weather or fluid flows) from sparse, irregular observations by partitioning them into "context" and "query" subsets.
- Ghost Imaging: Joost Batenburg developed Noise2Ghost, which uses random subsets of noisy measurements to reconstruct high-quality X-ray images without clean references.
- Connectomics: In X-ray neuronal microscopy, Alfred Laugros showed that splitting a single set of tomographic projections into two independent volumes allows for self-supervised denoising, enabling the identification of individual synapses in the brain.
Zero-Shot Learning: Training on a Single Patient
Perhaps the most radical application of this idea is Zero-Shot Self-Supervised Learning (ZS-SSL). As Burhaneddin Yaman describes in Zero-Shot Self-Supervised Learning for MRI Reconstruction, we can partition the data from a single scan into training, loss, and validation sets.
This allows a model to be optimized specifically for one patient's anatomy and one specific scanner's noise profile. By moving away from massive, generalized databases and toward patient-specific optimization, measurement splitting makes high-fidelity reconstruction possible even for rare pathologies or unique imaging conditions where no prior data exists.