---
title: 'iLAMA-Net: Initialization-Augmented SVCT Reconstruction'
url: https://www.emergentmind.com/topics/ilama-net
type: topic
---

# iLAMA-Net: Initialization-Augmented SVCT Reconstruction

iLAMA-Net is an initialization-augmented version of LAMA-Net for sparse-view CT (SVCT), introduced within the dual-domain variational framework of “LAMA-Net: A Convergent Network Architecture for Dual-Domain Reconstruction” [2507.22316]. It combines a geometry-aware initialization network, termed Init-Net, with the learned alternating minimization algorithm (LAMA) unrolled as LAMA-Net. The underlying model jointly reconstructs an image variable $x \in \mathbb{R}^n$ and a measurement-domain variable $z \in \mathbb{R}^m$ from an observed sparse sinogram $s_0 \in \mathbb{R}^{n_0}$, using the forward operator $A \in \mathbb{R}^{m \times n}$ and a view-selection mask $P_0 \in \mathbb{R}^{n_0 \times m}$, with fidelity terms that are quadratic and therefore correspond to an additive Gaussian-noise assumption, or a least-squares relaxation, in both image-to-sinogram consistency and measured-view consistency [2507.22316].

## 1. Dual-domain formulation and reconstruction objective

The reconstruction problem is posed as a two-block, nonconvex, nonsmooth optimization problem:
$$
\min_{x,z} \ \Phi(x,z; s_0, \Theta) := f(x,z; s_0) + R(x; \theta_1) + Q(z; \theta_2),
$$
with
$$
f(x,z; s_0) = \frac{1}{2}\|Ax-z\|_2^2 + \frac{\lambda}{2}\|P_0 z - s_0\|_2^2.
$$
Here $\Theta=(\theta_1,\theta_2)$ are learned parameters and $\lambda>0$ is a trade-off weight [2507.22316].

The learned regularizers are defined in the image and measurement domains as
$$
R(x; \theta_1) = \|g_R(x; \theta_1)\|_{2,1} = \sum_i \|g_{R,i}(x; \theta_1)\|_2,
$$
and
$$
Q(z; \theta_2) = \|g_Q(z; \theta_2)\|_{2,1} = \sum_j \|g_{Q,j}(z; \theta_2)\|_2.
$$
The feature extractors $g_R$ and $g_Q$ are lightweight CNNs of the form
$$
g^r(y) = w_\ell * a(\dots a(w_2 * a(w_1 * y)) \dots), \quad r \in \{R,Q\},
$$
where $a(\cdot)$ is a smoothed ReLU, “$*$” denotes convolution, and $\Theta$ collects these CNN weights [2507.22316].

The rationale for the dual-domain model is explicit. Sparse-view CT suffers from strong streaking and aliasing that are structured in both the sinogram and image domains. Enforcing $Ax \approx z$ and $P_0 z \approx s_0$ uses physics on both domains, while learning complementary priors $g_R$ and $g_Q$ allows the model to denoise and de-alias in the sinogram, enforce image priors attuned to anatomy, and couple both through the data-consistency term $f$. The paper states that this dual coupling is empirically superior to single-domain approaches for aggressive undersampling [2507.22316].

## 2. LAMA and the unrolled LAMA-Net architecture

To handle the nonsmooth $\ell_{2,1}$ regularizers, LAMA adopts a smoothed surrogate
$$
r_\epsilon(y)=\sum_i r_{\epsilon,i}(y),
$$
where
$$
r_{\epsilon,i}(y)=
\begin{cases}
\frac{1}{2\epsilon}\|g_i(y)\|_2^2, & \text{if } \|g_i(y)\|_2 \le \epsilon,\\
\|g_i(y)\|_2 - \frac{\epsilon}{2}, & \text{if } \|g_i(y)\|_2 > \epsilon,
\end{cases}
$$
for $r \in \{R,Q\}$. Its gradient is
$$
\nabla r_\epsilon(y)
=
\sum_{i\in I_0}\nabla g_i(y)^T \frac{g_i(y)}{\epsilon}
+
\sum_{i\in I_1}\nabla g_i(y)^T \frac{g_i(y)}{\|g_i(y)\|_2},
$$
with $I_0=\{i:\|g_i(y)\|_2 \le \epsilon\}$ and $I_1$ its complement [2507.22316].

For fixed $\epsilon$, LAMA uses a PALM-like alternating scheme with linearized proximal steps. In the measurement block,
$$
b_{k+1}=z_k-\alpha_k \nabla_z f(x_k,z_k),
$$
followed by the linearized update
$$
u^z_{k+1}=b_{k+1}-\hat{\alpha}_k \nabla Q_{\epsilon_k}(b_{k+1}),
\quad
\hat{\alpha}_k=\frac{\alpha_k p_k}{\alpha_k+p_k}, \ p_k>0.
$$
In the image block,
$$
c_{k+1}=x_k-\beta_k \nabla_x f(x_k,u^z_{k+1}),
$$
followed by
$$
u^x_{k+1}=c_{k+1}-\hat{\beta}_k \nabla R_{\epsilon_k}(c_{k+1}),
\quad
\hat{\beta}_k=\frac{\beta_k q_k}{\beta_k+q_k}, \ q_k>0.
$$
The gradients of the fidelity term are
$$
\nabla_x f(x,z)=A^T(Ax-z), \qquad
\nabla_z f(x,z)=(z-Ax)+\lambda P_0^T(P_0 z-s_0).
$$
These expressions define the data-consistency layers in the unrolled architecture [2507.22316].

LAMA includes sufficient descent checks. If the pair $(u^x_{k+1},u^z_{k+1})$ satisfies the stated sufficient descent conditions for some $\eta>0$, it is accepted. Otherwise, the algorithm falls back to a gradient BCD step with line search,
$$
v^z_{k+1}=z_k-\bar{\alpha}\,[\nabla_z f(x_k,z_k)+\nabla Q_{\epsilon_k}(z_k)],
$$
$$
v^x_{k+1}=x_k-\bar{\beta}\,[\nabla_x f(x_k,v^z_{k+1})+\nabla R_{\epsilon_k}(x_k)],
$$
with repeated reduction $(\bar{\alpha},\bar{\beta}) \leftarrow \rho(\bar{\alpha},\bar{\beta})$, $0<\rho<1$, until sufficient decrease is obtained [2507.22316].

LAMA-Net is obtained by unrolling this algorithm: each algorithm iteration is one network “phase,” and $K$ phases, for example $K=15$, form the network. The learned proximal terms are realized as residual corrections in both domains, while the fallback BCD descent with line search is implemented as a safeguard module. The paper distinguishes a U-module, corresponding to the linearized proximal residual steps, and a V-module, corresponding to the safeguarded fallback. Parameter sharing across all $K$ phases is used for memory efficiency and better generalization. Architectural details are explicit: $g_R$ has 4 convolutional layers, 32 channels, kernel $3 \times 3$, stride 1, activation $a(\cdot)$, and padding to preserve size; $g_Q$ has 4 convolutional layers, 32 channels, rectangular kernel $3 \times 15$ with padding $1 \times 7$, stride 1, and activation $a(\cdot)$. The use of rectangular kernels in the sinogram domain is stated to help capture along-view correlations aligned with CT geometry [2507.22316].

## 3. iLAMA-Net: initialization network and recurrent sinogram completion

iLAMA-Net augments LAMA-Net with an initialization network that produces informative $(x_0,z_0)$ from the sparse sinogram $s_0$. The motivation is that the objective is nonconvex and nonsmooth, and the paper states that better initialization substantially improves convergence speed, stability, and final accuracy [2507.22316].

The initialization mechanism is geometry-aware. It learns a recurrent mapping $\Psi_\omega$ that advances sparse-view sinogram content by one angular step $\Delta \theta$:
$$
s_1 \approx \Psi_\omega(s_0), \quad
s_2 \approx \Psi_\omega(s_1), \ \dots, \
s_{p-1} \approx \Psi_\omega(s_{p-2}).
$$
Here $p$ is the downsampling factor, for example $p=16$ for $1024 \rightarrow 64$ views, and $s_0,\dots,s_{p-1}$ are concatenated to approximate a pseudo full-view sinogram [2507.22316].

Training of $\Psi_\omega$ is performed by minimizing
$$
\min_\omega \frac{1}{pN}\sum_{n=1}^N \sum_{i=1}^p
\|\Psi_\omega(s^{(n)}_{i-1}) - s^{(n)}_i\|_2^2.
$$
After training, initialization is formed as
$$
z_\omega(s_0) = [s_0,\Psi_\omega(s_0),\dots,\Psi_\omega^{(p-1)}(s_0)],
$$
and
$$
x_0=\bar{x}_\omega(s_0):=\mathrm{FBP}(z_\omega(s_0)), \qquad z_0=z_\omega(s_0).
$$
The LAMA iteration itself is unchanged; it simply starts from these learned initials:
$$
(x_K,z_K)=\mathrm{LAMA}(x_0=\bar{x}_\omega(s_0), z_0=z_\omega(s_0)).
$$
This means that iLAMA-Net is not a new iterative rule but LAMA-Net supplied with a learned initialization pathway [2507.22316].

Init-Net is a CNN composed of 3 sequential blocks with skip connections; each block has 4 convolutional layers with $3 \times 15$ kernels, stride 1, padding $1 \times 7$, and ReLU activations. The paper states that these rectangular kernels align with sinogram geometry and capture along-view contextual correlations. Training uses ADAM for 100 epochs with learning rate $10^{-4}$ on full sinograms split into $p$ partitions. Init-Net and LAMA-Net are trained independently, which the paper describes as avoiding leakage [2507.22316].

The paper also gives an interpretation of why the initialization matters. Theoretically, a better $(x_0,z_0)$ is described as closer to feasible and high-quality regions, reducing the number of fallback steps, accelerating descent of $\Phi_\epsilon$, and promoting rapid satisfaction of the gradient reduction criterion that triggers $\epsilon$-decrease. Empirically, the recurrent sinogram completion is stated to capture angular correlations missed by one-shot inpainting, producing cleaner $x_0$ through FBP of $z_0$, improving final PSNR and SSIM, and further reducing sinogram RMSE [2507.22316].

## 4. Convergence, Clarke stationarity, and stability claims

The paper provides a complete and rigorous convergence proof for LAMA. The relevant nonsmooth notion is the Clarke subdifferential $\partial_c f(x,z)$ of a locally Lipschitz function, defined via the Clarke directional derivative
$$
f^\circ(x,z;r,s)=
\limsup_{(c,d)\to(x,z),\, t\downarrow 0}
\frac{f(c+tr,d+ts)-f(c,d)}{t},
$$
with
$$
\partial_c f(x,z)=\{(p,q): p^T r + q^T s \le f^\circ(x,z;r,s) \text{ for all } (r,s)\}.
$$
A Clarke stationary point satisfies $0 \in \partial_c f(x,z)$ [2507.22316].

The main convergence theorem is formulated for the subsequence of iterations at which the reduction criterion is satisfied and $\epsilon$ is reduced. Let $\{(x_k,z_k)\}$ be generated by LAMA with arbitrary $(x_0,z_0)$, $\epsilon_0>0$, and $\epsilon_{\mathrm{tol}}=0$. Let $\{(x_{k_\ell+1},z_{k_\ell+1})\}$ be the subsequence at indices $k=k_\ell$ where the gradient threshold is satisfied and $\epsilon$ is reduced. Then the subsequence has at least one accumulation point, and every accumulation point $(\bar{x},\bar{z})$ is a Clarke stationary point of the original unsmoothed objective $\Phi(x,z)=f(x,z)+R(x)+Q(z)$ [2507.22316].

The proof sketch reported in the paper rests on several ingredients: with $\epsilon$ fixed, the inner loop yields sufficient decrease and bounded gradient norm; Lipschitz continuity of $\nabla \Phi_\epsilon$ holds with $L_\epsilon=O(\epsilon^{-1})$; the line-search fallback guarantees finite termination at each iteration; the sequence is bounded by coercivity of $\Phi$; along the reduction subsequence, $\|\nabla \Phi_{\epsilon_k}(x_{k+1},z_{k+1})\|\to 0$ as $\epsilon_k \to 0$; and closedness of $\partial_c \Phi$, together with the explicit forms of $\nabla R_\epsilon$ and $\nabla Q_\epsilon$, yields Clarke stationarity of accumulation points [2507.22316].

A common misunderstanding is that this theorem implies global optimality. The paper explicitly frames the result as convergence of a reduction subsequence to a Clarke stationary point, which is weaker than global optimality. It also depends on the sufficient descent checks and the $\epsilon$-schedule. This suggests that the theoretical guarantee is best understood as a stationarity and descent guarantee for a nonconvex learned variational method, rather than a claim of uniqueness or global minimization [2507.22316].

The paper further links convergence behavior to stability and robustness. Because each accepted step reduces $\Phi_\epsilon$ or is safeguarded by a line search, the reconstruction is described as avoiding uncontrolled amplifications often seen in unconstrained unrolled networks. This is used to explain the observed robustness to structured perturbations and noise [2507.22316].

## 5. Experimental protocol and quantitative performance in sparse-view CT

The reported experiments use the AAPM-Mayo Clinic Low-Dose CT Grand Challenge dataset and NBIA. The training and test split is formed by randomly selecting 500 image-sinogram pairs from AAPM-Mayo and 200 from NBIA, then splitting each set 80%/20%. The acquisition geometry uses 512 detector elements and 1024 full views, with sparse-view settings of 64 views (6.25%) and 128 views (12.5%). Images are of size $256 \times 256$. The forward and backprojection operators are fan-beam distance-driven projector/backprojector from CTLIB. Sparse-view generation applies $P_0$ to the full sinogram $s$ to obtain zero-filled sparse $s_0$, and $\mathrm{FBP}(s_0)$ is used as a baseline and sometimes as $x_0$ [2507.22316].

LAMA-Net is trained with
$$
L(\Theta)=\frac{1}{N}\sum_i \Big[
\|x_K^{(i)}-\hat{x}^{(i)}\|_2^2
+
\|z_K^{(i)}-A\hat{x}^{(i)}\|_2^2
+
\mu \big(1-\mathrm{SSIM}(x_K^{(i)},\hat{x}^{(i)})\big)
\Big],
$$
with $\mu=0.01$ and $K=15$ phases. Optimization uses ADAM with learning rates $10^{-4}$ for the image network and $6 \times 10^{-5}$ for the sinogram network. Training is recursive: 3 phases for 300 epochs, then adding 2 phases and fine-tuning for 200 epochs, up to 15 phases. For iLAMA-Net, $\Psi_\omega$ is trained with the recurrence loss for 100 epochs using ADAM at learning rate $10^{-4}$, then $(x_0,z_0)$ are generated and LAMA-Net is trained on those initials [2507.22316].

| Dataset / views | LAMA-Net | iLAMA-Net |
|---|---:|---:|
| AAPM-Mayo, 64 views | 44.58 dB, SSIM 0.986, RMSE 0.68 $(\times 10^{-3})$ | 46.37 dB, SSIM 0.990, RMSE 0.52 $(\times 10^{-3})$ |
| AAPM-Mayo, 128 views | 50.01 dB, SSIM 0.995, RMSE 0.32 $(\times 10^{-3})$ | 51.02 dB, SSIM 0.996, RMSE 0.27 $(\times 10^{-3})$ |
| NBIA, 64 views | 41.40 dB, SSIM 0.976, RMSE 0.99 $(\times 10^{-3})$ | 42.11 dB, SSIM 0.979, RMSE 0.92 $(\times 10^{-3})$ |
| NBIA, 128 views | 45.20 dB, SSIM 0.988, RMSE 0.55 $(\times 10^{-3})$ | 47.28 dB, SSIM 0.992, RMSE 0.44 $(\times 10^{-3})$ |

Against the named baselines, the paper reports the following values. On AAPM-Mayo with 64 views: FBP 27.17 dB / 0.596, DDNet 35.70 dB / 0.923, LDA 37.16 dB / 0.932, DuDoTrans 37.90 dB / 0.952, LEARN++ 43.02 dB / 0.980, LAMA-Net 44.58 dB / 0.986, and iLAMA-Net 46.37 dB / 0.990. On AAPM-Mayo with 128 views: FBP 33.28 dB / 0.759, DDNet 42.73 dB / 0.974, LDA 43.00 dB / 0.976, DuDoTrans 43.48 dB / 0.985, LEARN++ 49.77 dB / 0.995, LAMA-Net 50.01 dB / 0.995, and iLAMA-Net 51.02 dB / 0.996. On NBIA with 64 views: FBP 25.72 dB / 0.592, DDNet 35.59 dB / 0.920, LDA 34.31 dB / 0.896, DuDoTrans 35.53 dB / 0.938, LEARN++ 38.53 dB / 0.956, LAMA-Net 41.40 dB / 0.976, and iLAMA-Net 42.11 dB / 0.979. On NBIA with 128 views: FBP 31.86 dB / 0.743, DDNet 40.23 dB / 0.961, LDA 40.26 dB / 0.963, DuDoTrans 40.67 dB / 0.976, LEARN++ 43.35 dB / 0.983, LAMA-Net 45.20 dB / 0.988, and iLAMA-Net 47.28 dB / 0.992. In all four cases, the paper also reports lower sinogram RMSE for iLAMA-Net than for LAMA-Net [2507.22316].

The initialization ablation on AAPM-Mayo isolates the contribution of Init-Net. Without initialization, using FBP as $x_0$ and $s_0$ as $z_0$, LAMA-Net attains 44.58/0.986 at 64 views and 50.01/0.995 at 128 views. With a CNN inpainting initialization in the image domain only, the corresponding values are approximately 45.1/0.987 and approximately 50.7/0.995. With the geometry-aware sinogram recurrence, iLAMA-Net attains the best values, 46.37/0.990 and 51.02/0.996 [2507.22316].

## 6. Model size, robustness, limitations, and practical implications

The reported parameter counts are: DDNet $6 \times 10^5$, LDA $6 \times 10^4$, DuDoTrans $8 \times 10^6$, LEARN++ $6 \times 10^6$, LAMA-Net $3 \times 10^5$, and iLAMA-Net $4 \times 10^5$. The paper states that LAMA-Net and iLAMA-Net achieve SOTA accuracy with orders-of-magnitude fewer parameters than several baselines. A plausible implication is that the compactness follows from the use of small CNN regularizers and parameter sharing across phases, rather than transformer-scale representational capacity [2507.22316].

Qualitatively, LAMA-Net and iLAMA-Net are reported to suppress streaking while preserving fine edges and textures, and the sinogram RMSE improvements are taken as confirmation of better measurement-domain consistency. The paper also presents stability experiments. Under a structured perturbation in which “can u see it” text is injected into the ground truth, LAMA-Net reconstructs the perturbation faithfully without hallucinations elsewhere, and the difference heatmaps largely show the perturbation only. Under Gaussian perturbations, LAMA-Net consistently attains higher PSNRs against perturbed ground truths across several noise levels, and its difference images exhibit mostly isotropic, structureless noise rather than spurious edges [2507.22316].

The dual-domain coupling is further described as making the method less brittle as the number of views decreases, because $Ax \approx z$ and $P_0 z \approx s_0$ constrain both domains while $g_R$ and $g_Q$ capture cross-domain priors. The ablation study indicates that iLAMA-Net further improves robustness by preventing bad local minima and reducing artifact propagation from extreme undersampling. This suggests that initialization is not merely a performance enhancement but also part of the method’s robustness profile [2507.22316].

Several limitations are explicit. The method requires accurate $A$ and $P_0$, so mismatch across scanners or geometries can degrade performance. The convergence theorem concerns a reduction subsequence and Clarke stationary points, not global optima. Nonconvexity remains fundamental because the learned priors make $\Phi$ nonconvex and allow multiple stationary points. The paper also notes that extreme out-of-distribution cases still pose challenges. Proposed future directions include multi-domain priors such as Fourier or wavefront domains, physics-informed constraints such as dose models, uncertainty quantification via Bayesian or ensemble approaches atop the convergent backbone, and joint learning of $A$-calibration and priors for cross-scanner portability [2507.22316].

From an implementation standpoint, the paper gives a concrete inference procedure: initialize $z_0$ with recurrent sinogram completion, set $x_0=\mathrm{FBP}(z_0)$, perform $K$ LAMA phases with the same accept-or-reject rules and optional $\epsilon$ update, and return $x_K$. It states that $K=15$ phases balances accuracy and speed; $\alpha_k$, $\beta_k$, $\hat{\alpha}_k$, and $\hat{\beta}_k$ can be learned as trainable scalars per phase; safeguard parameters typically use $\delta,\eta \in (0,1)$, $\rho \in (0,1)$, $\gamma \in (0,1)$, and $\sigma>0$; and inference time depends on the projector implementation and GPU. The experiments were run on an NVIDIA A100 80GB [2507.22316].

Source: https://www.emergentmind.com/topics/ilama-net