Limited-Angle Tomography Reconstruction via Projector Guided 3D Diffusion

TL;DR

Limited-angle electron tomography suffers from missing-wedge artifacts that hinder accurate 3D reconstruction. TEMDiff introduces a projector-guided 3D diffusion framework trained on FIB-SEM data through a physics-based simulator, enforcing cross-slice consistency and data fidelity at each denoising step, and it generalizes to real TEM tilts down to $8^ leftarrow$ with $2^ ightarrow$ increments without retraining. It achieves state-of-the-art or near-state-of-the-art performance on synthetic datasets and real tilts, with pronounced gains at extreme angles ($\leq 10^ ext{o}$), demonstrating robust 3D coherence and artifact suppression. By transferring structural priors from abundant FIB-SEM data and integrating a projection-based correction, TEMDiff broadens the applicability of LACT in biology and materials science under severe angular constraints.

Abstract

Limited-angle electron tomography aims to reconstruct 3D shapes from 2D projections of Transmission Electron Microscopy (TEM) within a restricted range and number of tilting angles, but it suffers from the missing-wedge problem that causes severe reconstruction artifacts. Deep learning approaches have shown promising results in alleviating these artifacts, yet they typically require large high-quality training datasets with known 3D ground truth which are difficult to obtain in electron microscopy. To address these challenges, we propose TEMDiff, a novel 3D diffusion-based iterative reconstruction framework. Our method is trained on readily available volumetric FIB-SEM data using a simulator that maps them to TEM tilt series, enabling the model to learn realistic structural priors without requiring clean TEM ground truth. By operating directly on 3D volumes, TEMDiff implicitly enforces consistency across slices without the need for additional regularization. On simulated electron tomography datasets with limited angular coverage, TEMDiff outperforms state-of-the-art methods in reconstruction quality. We further demonstrate that a trained TEMDiff model generalizes well to real-world TEM tilts obtained under different conditions and can recover accurate structures from tilt ranges as narrow as 8 degrees, with 2-degree increments, without any retraining or fine-tuning.

Figures (11)

• Figure 1: Mitochondria reconstructed using FBP, SART, DiffusionMBIR, DOLCE, and TEMDiff, with tilts covering $10^\circ$ with $1^\circ$ increments. The top row presents 3D views, while the bottom row displays vertical cross-section along x axis of each corresponding 3D view. TEMDiff produces more clear, more consistent and better structures than others in both views. The contrast for these other methods has been enhanced for better visual quality.
• Figure 2: The TEMDiff pipeline. (Top) At inference time, given the acquired tilts $\mathbf{y}$, the inverse radon transform $C=\mathcal{R}^{-1}(\mathbf{y})$ is computed and concatenated with noise. The result is fed to the pretrained U-Net $\mathcal{N}$ which includes attention layers and takes the acquisition angular range $\theta$ and increments $\Delta\theta$ as further inputs. The U-Net $\mathcal{N}$ estimates the amount of noise in the input and iteratively denoise it. (Bottom) To train $\mathcal{N}$ without access to ground-truth data, FIB-SEM data is leveraged to synthesize realistic STEM tilts $\Hat{\mathbf{y}}$, given $\theta$ and $\Delta\theta$. Then, the inverse radon transform $C=\mathcal{R}^{-1}(\Hat{\mathbf{y}})$ is computed and concatenated with the noised ground truth patch. The result is fed to $\mathcal{N}$ together with $\theta$ and $\Delta\theta$ to predict the noise in the input, which is used to train $\mathcal{N}$ by minimizing the mean squared error.
• Figure 3: Reconstructions of FBP, SART and TEMDiff with real tilts of $8^\circ$ angular range and $1^\circ$ or $2^\circ$ increments. The reference volumes are reconstructed using AreTomo with all available tilts ($80^\circ$ or $120^\circ$). Each main figure is the 3D view of reconstruction and its bottom-right inset corresponds to one slice of 2D view. TEMDiff produces more realistic and clearer reconstructions with good contrast on both 3D and 2D views. In 3D views, the contrast of FBP and SART are enhanced for better visual quality.
• Figure 4: Samples of different datasets. FIB-SEM image of (a) mitochondria in brain cells. (b) and (c) mitochondria in two different HeLa cells. (d) synapse in brain cells FIB-SEM. Reconstruction using AreTomo from real TEM tilts of (e) and (f) mitochondria, (g) and (h) synapse.
• Figure 5: Histogram (left) and examples (right) of (a) real STEM tilts, (b) simulated STEM tilts with proposed mapping, and (c) radon transform of FIB-SEM volume. The tilts synthesized by the proposed mapping are visually and statistically close to real STEM tilts.