3D-FastResUNet: Efficient 3D Reconstruction

• 3D-FastResUNet is a lightweight, efficient 3D convolutional neural network that integrates residual learning, squeeze-and-excitation modules, and depthwise separable convolutions for dynamic image reconstruction.
• It employs an encoder-decoder design with max-pooling, ASPP, and attention-based skip fusion to capture multiscale context and enhance feature refinement.
• The architecture significantly reduces convergence iterations and computational load, making it well-suited for fast, unsupervised tomographic applications such as pulmonary impedance imaging.

3D-FastResUNet is a lightweight, highly efficient 3D convolutional neural network architecture specifically designed to serve as the reconstruction backbone for Deep Dynamic Image Prior (D²IP) in 3D time-sequence tomographic applications such as pulmonary impedance imaging. It is characterized by a reductionist adaptation of the 3D U-Net paradigm, utilizing residual connections, squeeze-and-excitation (SE) mechanisms, attention-based skip fusion, and depthwise separable convolutions, all tailored for robust performance in training-data-free, unsupervised image prior methods. The architecture is engineered to address the significant computational costs associated with classic Deep Image Prior (DIP) approaches in 3D and dynamic inverse imaging scenarios, achieving rapid convergence and low memory usage without sacrificing reconstruction quality (Fang et al., 18 Jul 2025).

1. Architectural Composition

3D-FastResUNet adopts an encoder–decoder topology with three downsampling and upsampling stages, a streamlined alternative to deeper variants. The encoder begins with a stem block: a 3D convolution of kernel size $3^3$ (stride=1, padding=1), LayerNorm, and LeakyReLU, producing 32 channels. Three successive downsampling stages sequentially increase channel depth (32→64, 64→128, 128→256) through ResConv3D blocks. Each ResConv3D block consists of two 3D convolutional layers with pre-activation residual connections, LayerNorm, and LeakyReLU; SE modules are inserted after each block to recalibrate channel-wise features. Downsampling is effected by $2^3$ max-pooling.

At the bottleneck, atrous spatial pyramid pooling (ASPP) with dilation rates {1, 6, 12, 18} (all at 256 channels) is applied to capture multiscale context. The decoder mirrors the encoder, using trilinear interpolations (×2) for upsampling, and introduces attention gates for skip feature fusion at each stage. Decoder ResConv3D blocks reverse the channel scaling (256→128, 128→64, 64→32). The final tail block uses a $1^3$ convolution to map 32 channels to 1, followed by a sigmoid activation.

Stage	Operation	Output Channels
Stem	Conv3D + LayerNorm + LeakyReLU	32
Downsample 1	ResConv3D + SE + MaxPool	64
Downsample 2	ResConv3D + SE + MaxPool	128
Downsample 3	ResConv3D + SE + ASPP	256
Upsample 1	Interp×2 + Attention + ResConv3D	128
Upsample 2	Interp×2 + Attention + ResConv3D	64
Upsample 3	Interp×2 + Attention + ResConv3D	32
Tail	Conv3D (1^3) + Sigmoid	1

2. Residual Block and Attention Mechanisms

Each ResConv3D employs a pre-activation design. Formally, for input $x \in \mathbb{R}^{C \times D \times H \times W}$, weights $W_1$, $W_2$, and LayerNorm $\mathcal{L}(\cdot)$, LeakyReLU $\sigma(\cdot)$, the output is: $$F(x) = W_2 * \sigma(\mathcal{L}(W_1 * x)), \quad y = x + F(x).$$ The calculation of core 3D convolution: $$[x * W]_{n,c_\mathrm{out},i,j,k} = \sum_{c_\mathrm{in}=1}^{C_\mathrm{in}} \sum_{u=-\lfloor K/2 \rfloor}^{\lfloor K/2 \rfloor} \sum_{v=-\lfloor K/2 \rfloor}^{\lfloor K/2 \rfloor} \sum_{w=-\lfloor K/2 \rfloor}^{\lfloor K/2 \rfloor} x_{n,c_\mathrm{in},i+u,j+v,k+w} W_{c_\mathrm{out},c_\mathrm{in},u,v,w}.$$ Decoder-stage attention gates combine skip and upsampled features via: $2^3$0

3. Distinctive Design Choices

3D-FastResUNet diverges from standard 3D U-Net architectures in several key dimensions:

• The depth of down/up-sampling is reduced to three stages, as opposed to the 4–5 found in typical implementations.
• Residual blocks with LayerNorm and LeakyReLU substitute for conventional conv–BN–ReLU sequences, enhancing gradient propagation and mitigating vanishing/exploding gradient issues for deeper representational power.
• Depthwise separable 3D convolutions (Xception-style) are utilized, which result in approximately 40% fewer floating-point operations (FLOPs) and model parameters.
• SE units are inserted within the encoder to selectively amplify informative feature channels.
• ASPP at the bottleneck augments the receptive field without necessitating further spatial downsampling.
• Decoder skip connection fusion employs attention gates instead of simple concatenation and convolution, focusing reconstruction on salient features.
• LayerNorm is used throughout, rather than BatchNorm, to optimize for single-instance (one-case-at-a-time) DIP-style training.

These choices collectively realize fast convergence—typical DIP application converges in 200–250 iterations per frame (instead of >1000), with substantial benefits in computational and memory demands (Fang et al., 18 Jul 2025).

4. Network Hyperparameters and Optimization

The entire 3D-FastResUNet comprises approximately 3.2 million parameters. Depthwise separable convolutions critically contribute to this compactness. The model uses LeakyReLU activations with a negative slope of 0.2 for all internal layers and a sigmoid at the output. Convolutional kernels are $2^3$1 throughout (except for the final $2^3$2 reduction layer). Kaiming (He) initialization is applied to all convolutional weights.

Training employs the Adam optimizer with a learning rate of $2^3$3 for simulation data, or $2^3$4 for clinical scenarios. Regularization includes a 4D total variation (TV) term added to the DIP loss, with $2^3$5, $2^3$6, and overall $2^3$7.

5. Performance Metrics and Ablation Findings

Empirical results on 3D time-sequence Electrical Impedance Tomography (tsEIT) data demonstrate significant performance advantages in both speed and accuracy:

Method	Inference Time (20 frames, s)	Memory (GB, RTX 4090)	Convergence Iterations
3D-FastResUNet (D²IP)	135 (6.8 per frame)	4.3	200–250
R-SIP (SOTA unsupervised)	960 (48 per frame)	—	>1000
TV-regularized	648 (32.4 per frame)	—	—
Tikhonov	595 (29.8 per frame)	—	—

Ablation experiments reveal that omitting Unsupervised Parameter Warm-Start (UPWS) and/or Temporal Parameter Propagation (TPP) in D²IP substantially increases runtime (~1200s and ~450s per 20-frame sequence, respectively), confirming that both strategies contribute to 70%+ runtime reduction. The aggregate effect accelerates per-frame convergence from ~1000 to <300 iterations (Fang et al., 18 Jul 2025).

6. Application Context and Suitability

3D-FastResUNet is specifically validated within D²IP as an implicit image prior for unsupervised 3D time-sequence imaging. In this regime, it addresses the high computational overhead and extended optimization typical of DIP-based volumetric and dynamic reconstructions. Incorporation of depthwise separable convolutions, attention modules, and LayerNorm yields a model capable of stable single-instance training, rapid convergence, and reduced hardware requirements—crucial for clinical scenarios such as dynamic pulmonary imaging.

A plausible implication is that the architecture’s modularity and general efficiency may make it transferable to other volumetric tomographic tasks requiring unsupervised or weakly supervised priors, particularly where memory constraints and rapid inference are pivotal (Fang et al., 18 Jul 2025).