---
kind: concept
slug: data-consistency-layer
title: Data-consistency layer
summary: >-
  A step inside learned image reconstruction that pulls the output back toward the measured samples. As a layer, it sets sampled k-space entries to the measurements (noise-free) or mixes them with a weight λ (noisy). As a GAN loss, a cyclic term plays the same role. Shown for cardiac and compressed-sensing MRI.
tags: [mri-reconstruction, inverse-problems, data-fidelity, medical-imaging, cyclic-loss]
papers: ["1704.02422", "1709.00753"]
related: [deep-unfolding, convolutional-neural-network, generative-adversarial-network, generative-prior-inverse-problems]
status: draft
updated_at: 2026-10-19
---

## Overview

A learned reconstruction gets a set of measurements y and must output an image
that agrees with them. A network trained alone can drift away from the
measurements, because nothing in its forward pass forces agreement with the
data. The data-consistency (DC) layer is a fixed, differentiable step placed
after each network block. It transforms the output into k-space and combines
it with the entries that were actually measured. The invariant is that the
measured entries of the output equal the measurements, or sit close to them
when the measurements are noisy [arxiv:1704.02422 §2] [arxiv:1704.02422 §3].

In the paper's setting, k-space is the Fourier transform of the image, and the
measurements are a masked subset of it. The paper argues that a plain CNN from
the aliased image ignores the measured data unless the data term is part of
the learning [arxiv:1704.02422 §2]. The DC layer supplies that term without
adding learned parameters in the noise-free case [arxiv:1704.02422 §3].

A second family enforces the same idea through the loss instead of the
architecture. In a GAN-based reconstruction, the adversarial loss alone does
not force the generator to respect the measurements, so a cyclic term is added
that compares the generator's output with them [arxiv:1709.00753 §III-D].

## Variants

### Noise-free hard replacement

On the sampled set Ω, the layer sets the k-space value to the measured value.
Off Ω it keeps the network output. This is the limit of the weighted form as
the weight λ goes to infinity [arxiv:1704.02422 §3]. The paper uses this limit
for its noise-free experiments [arxiv:1704.02422 §3]. Card:
[Schlemper et al. 2017, cascaded CNN for dynamic cardiac MRI](/papers/1704.02422).

### Noisy soft data consistency

When the measurements carry noise, the paper uses a finite weight λ. On Ω the
output is a convex combination of the network value and the measurement, with
weights 1/(1+λ) and λ/(1+λ) [arxiv:1704.02422 §3, eq. 6]. The paper sets
λ = q/σ, where σ is the noise level, starts λ at 0.025, and makes it trainable
[arxiv:1704.02422 §3]. In its noise experiments λ ends near 0.5 after
fine-tuning [arxiv:1704.02422 §7.3.3]. The paper gives the training range of σ²
but no numbers for q or σ [arxiv:1704.02422 §7.3.3].

### Loss-based data consistency (cyclic loss in a GAN)

A GAN generator is trained with an adversarial loss plus a cyclic term. The
cyclic term has two parts. The first takes the generator's output for the
measured samples, under-samples it again with the same mask, and compares the
result with the measurements. This is the data-consistency part. The second
applies the same round trip to fully sampled images [arxiv:1709.00753 §III-D,
eq. 6]. The distance d in the first part may be MSE, MAE or another pixel-wise
distance. The paper does not say which one its experiments use
[arxiv:1709.00753 §III-D]. The weights are fixed at α = 1.0 for the first part
and γ = 10.0 for the second, for all experiments [arxiv:1709.00753 §III-F,
eq. 7]. The authors report that the cyclic loss raised PSNR by about 1 dB over
the single-generator baseline. This is a text-only figure in the paper
[arxiv:1709.00753 §IV-A]. Card:
[Compressed Sensing MRI Reconstruction using a GAN with a Cyclic Loss](/papers/1709.00753).

### Cascade with data-sharing (dynamic sequences)

For a time series, the paper pairs the DC layer with a data-sharing layer. For
each frame it averages the sampled k-space of neighbouring frames in a window,
and fills only the unsampled locations with that average. The averages enter
the next block as extra input channels [arxiv:1704.02422 §5] [arxiv:1704.02422 §6].
The layer applies to sequences only, not to 2D frames [arxiv:1704.02422 §6].
The paper reports lower error with data sharing than without it at
9-fold undersampling in its 3D setting [arxiv:1704.02422 fig. 10]
[arxiv:1704.02422 §7.3.1]. The layer is a companion to the DC step, not a
replacement for it.

## When to use which

| Situation | Use | Why |
|---|---|---|
| Measurements are noise-free, or noise is negligible | Hard replacement (λ → ∞) | Measured entries are kept exactly. The paper's noise-free experiments use this limit [arxiv:1704.02422 §3]. |
| Measurements are noisy and the noise level can be estimated | Soft DC with weight λ | λ lets the output move away from the data where the network is more reliable. The paper's trained λ was near 0.5 after fine-tuning [arxiv:1704.02422 §7.3.3]. The paper gives no numbers for q or σ, so a reimplementation has to choose λ itself. |
| Learned model from an aliased or undersampled input | Add DC after each block | The paper argues a plain CNN ignores the measured k-space [arxiv:1704.02422 §2]. |
| Several cascade stages are affordable | More stages, each with a DC step | Error fell at each stage from n_c = 1 to 5, with smaller gains after n_c = 3 [arxiv:1704.02422 fig. 7] [arxiv:1704.02422 §7.2.2]. |
| Tight memory or training budget | Fewer stages, or patch-based training | The 2D cascade of five blocks used about 335 MB of training activations at 256 × 256 [arxiv:1704.02422 §7.4]. The 3D model trains on 256 × 32 × 30 patches, whose activations alone took about 2.4 GB [arxiv:1704.02422 §7.4]. |
| Frames in a sequence share structure | Add the data-sharing layer | It lowers error at 9-fold undersampling in the paper's ablation [arxiv:1704.02422 fig. 10]. The paper reports no comparison at other rates for this ablation. |
| Generator trained adversarially, with no fixed projection step | Cyclic loss term | The adversarial loss alone does not enforce agreement with the measurements [arxiv:1709.00753 §III-D]. A loss only pushes toward agreement, so it does not guarantee that measured entries are reproduced exactly. A hard or soft layer does. |

## Implementation sketch

The steps below are the paper's eq. 6 for a single 2D frame, the cascade that
uses it, and the cyclic loss term. Shapes and the FFT convention are in the
paper card's pseudocode.

```pseudocode
# Data-consistency step, paper eq. 6  [arxiv:1704.02422 §3]
# s_cnn: network output in k-space (complex)
# s_0:   measured k-space, zero off the sampled set
# omega: boolean mask, true where a sample was acquired
# lam:   weight; None means the noise-free limit (lambda -> infinity)

def data_consistency(s_cnn, s_0, omega, lam):
    # Start from the network output everywhere  [arxiv:1704.02422 §3]
    out = s_cnn.copy()
    if lam is None:
        # Noise-free: replace sampled entries with the measurements
        # [arxiv:1704.02422 §3]
        out[omega] = s_0[omega]
    else:
        # Noisy: weighted mix of output and measurement on Ω, divided
        # by (1 + lam)  [arxiv:1704.02422 §3, eq. 6]
        out[omega] = (s_cnn[omega] + lam * s_0[omega]) / (1 + lam)
    # Entries off Ω keep the network output  [arxiv:1704.02422 §3, eq. 6]
    return out
```

```pseudocode
# Cascade: block, then DC, repeated  [arxiv:1704.02422 §4] [§6]
# n_c stages; each block output is added to its input (residual)

def cascade(x_u, s_0, omega, n_c, lam):
    x = x_u
    for stage in range(n_c):
        # Residual CNN block  [arxiv:1704.02422 §6]
        x = x + cnn_block(x)
        # Back to k-space, apply DC, and back to the image domain
        # [arxiv:1704.02422 §3]
        s = fft2(x)
        s = data_consistency(s, s_0, omega, lam)
        x = ifft2(s)
    return x
```

```pseudocode
# Cyclic data-consistency term, first part of paper eq. 6  [arxiv:1709.00753 §III-D]
# m:    measured samples, under-sampled by the same mask
# dist: pixel-wise distance; the paper does not say which one it uses

def cyclic_dc_loss(gen, zero_filled, measured, mask, dist):
    # Generator output for the measured sample  [arxiv:1709.00753 §III-D]
    recon = gen(zero_filled)
    # Under-sample the output again with the same mask  [arxiv:1709.00753 §III-D]
    resampled = mask * fft2(recon)
    # Compare with the measurements that were actually taken  [arxiv:1709.00753 §III-D]
    # The caller adds this term with weight alpha = 1.0  [arxiv:1709.00753 §III-F, eq. 7]
    return dist(measured, resampled)
```

Notes for implementers:

- The layer is a fixed linear map on the output, so in the noise-free case it
  adds no trainable parameters [arxiv:1704.02422 §3]. In the noisy case the
  paper makes λ trainable [arxiv:1704.02422 §3]. Gradients flow through the
  layer, and the paper trains the whole cascade jointly
  [arxiv:1704.02422 §4].
- The paper's noisy-branch formula, eq. 6, leaves unacquired entries
  unchanged. The official code also scales those entries by 1/(1 + λ). The
  paper card's Paper vs code section records this. Implement the equation as
  written, or document the difference.
- The paper computes the forward model with a unitary 2D DFT per frame
  [arxiv:1704.02422 §2]. Centred and non-centred FFT conventions move the
  mask; keep the mask and the FFT in one convention.
- The cyclic term is a soft constraint, and its distance is not stated. Each
  reimplementation has to pick d and document the choice
  [arxiv:1709.00753 §III-D].

## Contradictions

Within 1704.02422, the text and the released code disagree on the noisy DC
branch and on whether λ is trainable in the released code. The paper card
records both points [arxiv:1704.02422 §3].

Between the two papers there is no direct contradiction, but their numbers do
not compare. 1704.02422 reports MSE on cardiac cine data [arxiv:1704.02422
§7.1]. 1709.00753 reports PSNR, SSIM and NRMSE on IXI brain and Data Science
Bowl chest images [arxiv:1709.00753 §IV-A]. Different data, metrics and
baselines make the two sets of results unrelated quantitatively.

## Papers

- [Schlemper et al. 2017, cascaded CNN for dynamic cardiac MRI](/papers/1704.02422): introduces the DC layer as a fixed step after each CNN block, its noise-free and noisy forms, and the data-sharing layer for dynamic cardiac MRI.
- [Compressed Sensing MRI Reconstruction using a GAN with a Cyclic Loss](/papers/1709.00753): uses a cyclic loss that ties a GAN generator's output back to the measurements, and chains a second generator for refinement.