Deconver: A Deconvolutional Network for Medical Image Segmentation

arXiv:2504.00302 · eess.IV, cs.CV · Submitted 2025-04-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Deconver: A Deconvolutional Network for Medical Image Segmentation".

Paper summary: Rosa: We've seen how Deconver integrates deconvolution techniques into a U-shaped architecture for medical image segmentation and it claims state-of-the-art performance on both two dimensional and three dimensional segmentation tasks across four different medical datasets.

Dev: It’s notable because they manage to do this with substantially fewer computational costs and parameters than the leading methods currently out there.

Rosa: The core idea is that Deconver replaces computationally expensive attention mechanisms with efficient nonnegative deconvolution operations, enabling the restoration of high-frequency details while suppressing artifacts.

Dev: That means they are using nonnegativity and deconvolution to cut down on the heavy math without sacrificing accuracy on complex scans.

Rosa: They also introduced a backpropagation friendly layer based on a provably monotonic update rule for nonnegative deconvolution, which is key for stable end-to-end training using current deep learning frameworks.

Dev: That stability from that update rule seems important because it makes the whole process work with current deep learning frameworks.

Rosa: The paper also shows how they achieved state of the art performance in Dice scores and Hausdorff distance across ISLES’twenty-two BraTS’twenty-three GlaS, and FIVES datasets.

Dev: So those are the specific metrics they used to prove superiority on both 2D and three dee tasks.

Rosa: By bridging classical image restoration with modern deep learning, Deconver provides a practical solution for high precision segmentation, particularly in resource constrained clinical workflows.

Dev: That’s the implication for real world use, making it more applicable where you can't always afford the biggest infrastructure.

Conclusion: Rosa: We're wrapping up our discussion on Deconver, which is this network that puts deconvolution techniques right inside a U-shaped structure for medical image segmentation.

Dev: It’s notable because they manage to do this with substantially fewer computational costs and fewer parameters than the leading methods currently out there.

Rosa: The main takeaway is that they replaced those expensive attention mechanisms with these efficient deconvolution operations which help restore fine details while cleaning up artifacts.

Rosa: Ultimately, this work provides a practical solution for high precision segmentation in resource constrained clinical workflows by bridging classical image restoration with modern deep learning.

Dev: So Deconver is showing us how we can get better segmentation performance while also cutting down on the processing load and parameter count compared to other leading methods.

Department of Electrical Engineering at STADIUS Center, KU Leuven, Belgium · Data Science for Health Unit at Fondazione Bruno Kessler, Trento, Italy · Center for Biomedical Informatics Research at Stanford University · Division of Hospital Medicine at Stanford University

eess.IV, cs.CV

Submitted: 2025-04-01

Updated: 2025-04-01

Comments: 12 pages, 6 figures, 5 tables

DOI: 10.1109/JBHI.2025.3623381

Code: https://github.com/pashtari/deconver

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 98/100

The gist: Deconver introduces a novel network that integrates traditional deconvolution techniques from image restoration as a core learnable component within a U-shaped architecture.

Key concepts

Deconv Mixer Module
This is a learnable component in Deconver that replaces the standard multi-head self-attention module. It is designed to integrate deconvolution principles directly into the network structure, allowing it to perform image restoration tasks while learning optimal parameters for segmentation.
Nonnegative Deconvolution (NDC)
NDC is a specific layer used in Deconver that operates by partitioning input features into groups and refining a source image using a multiplicative update rule. This process helps restore high-frequency details from the input data while ensuring the resulting features remain non-negative, which is crucial for accurate image reconstruction.
Multiplicative Update Rule
This mathematical operation is applied within the NDC layer to refine the source image during training. A key advantage of this rule is its proven ability to reduce reconstruction errors, ensuring that Deconver effectively restores necessary high-frequency details without introducing excessive artifacts into the final segmentation output.

Terminology

Summary

Deconver introduces a novel network that integrates traditional deconvolution techniques from image restoration as a core learnable component within a U-shaped architecture. p. 0

Where the paper says it

Chen , Giuseppe Jurman , Aleksandra Pizurica ˇ , Senior Member, IEEE, and Sabine Van Huffel Abstract— While convolutional neural networks (CNNs) and vision transformers (ViTs) have advanced medical image segmentation, they face inherent limitations such as local receptive fields in CNNs and high computational complexity in ViTs. This paper introduces Deconver, a novel network that integrates traditional deconvolution techniques from image restoration as a core learnable component within a U-shaped architecture.

Deconver achieves state-of-the-art performance on both 2D and 3D segmentation tasks with substantially fewer computational costs and parameters than leading baselines. p. 1
Where the paper says it

• Nonnegative deconvolution layer: We introduce a backpropagation-friendly, differentiable layer based on a provably monotonic update rule for nonnegative deconvolution, enabling stable end-to-end training using current deep learning frameworks. • Performance and efficiency: Deconver achieves stateof-the-art performance on both 2D and 3D segmentation tasks with substantially fewer computational costs and parameters than leading baselines.

p. 1
Where the paper says it

• Nonnegative deconvolution layer: We introduce a backpropagation-friendly, differentiable layer based on a provably monotonic update rule for nonnegative deconvolution, enabling stable end-to-end training using current deep learning frameworks. • Performance and efficiency: Deconver achieves stateof-the-art performance on both 2D and 3D segmentation tasks with substantially fewer computational costs and parameters than leading baselines.

By bridging traditional image restoration with modern deep learning, Deconver provides a practical solution for high-precision segmentation in resource-constrained clinical workflows. p. 1
Where the paper says it

Extensive experiments across four datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate Deconver’s superiority in Dice scores and boundary accuracy (Hausdorff distance). By bridging classical image restoration with modern deep learning, Deconver provides a practical solution for highprecision segmentation, particularly in resource-constrained clinical workflows.

p. 1
Where the paper says it

Extensive experiments across four datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate Deconver’s superiority in Dice scores and boundary accuracy (Hausdorff distance). By bridging classical image restoration with modern deep learning, Deconver provides a practical solution for highprecision segmentation, particularly in resource-constrained clinical workflows.

Architectural Innovation

Deconver is the first network to incorporate deconvolution principles as a learnable component within a deep architecture. Key innovations include replacing computationally expensive attention mechanisms with efficient nonnegative deconvolution (NDC) operations, enabling the restoration of high-frequency details while suppressing artifacts. p. 0

Where the paper says it

Deconver replaces computationally expensive attention mechanisms with efficient nonnegative deconvolution (NDC) operations, enabling the restoration of high-frequency details while suppressing artifacts. Key innovations include a backpropagation-friendly NDC layer based on a provably monotonic update rule and a parameter-efficient design.

The main contributions of this work are threefold: Architectural Innovation, Nonnegative deconvolution layer, and Performance and efficiency. p. 1
Where the paper says it

The main contributions of this work are threefold: • Architectural Innovation: Deconver is the first network to incorporate deconvolution principles as a learnable component within a deep architecture. • Nonnegative deconvolution layer: We introduce a backpropagation-friendly, differentiable layer based on a provably monotonic update rule for nonnegative deconvolution, enabling stable end-to-end training using current deep learning frameworks.

How the training signal is built

The Deconver block replaces the multi-head self-attention module with a learnable Deconv Mixer module and substitutes layer normalization with instance normalization. The NDC layer operates in a grouped manner, partitioning an input feature map X into G groups along the channel dimension. p. 4

Where the paper says it

Nonnegative Deconvolution Layer The nonnegative deconvolution (NDC) layer forms the core innovation of Deconver, incorporating nonnegative deconvolution (presented in Section III-B) as a learnable layer to enhance feature representations. The NDC layer operates in a grouped manner, partitioning an input feature map X ∈ R C×H×W×D ≥0 into G groups {X g} G g=1 along the channel dimension.

Each group X g is processed independently, maintaining its own learnable filter Vg ≥ 0 and source image Sg ≥ 0. p. 4
Where the paper says it

The NDC layer operates in a grouped manner, partitioning an input feature map X ∈ R C×H×W×D ≥0 into G groups {X g} G g=1 along the channel dimension. Each group X g ∈ R C/G×H×W×D ≥0 is processed independently, maintaining its own learnable filter Vg ≥ 0 and source image Sg ≥ 0.

The NDC layer applies a single iteration of the multiplicative update rule (3) to refine the source image. p. 4
Where the paper says it

The filter Vg ∈ R C/G×RC/G×(2M+1)×(2N+1) ≥0 is initialized using the Kaiming uniform distribution [27] and clamped to nonnegative values via ReLU before being plugged into the update rule. The NDC layer applies a single iteration of the multiplicative update rule (3) to refine the source image.

Where the gains come from

The key advantage of the multiplicative update (3) is its guarantee of the reduction of the reconstruction error, which is proven in theorem 1. Extensive experiments across four datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate Deconver’s superiority in Dice scores and boundary accuracy (Hausdorff distance). p. 1

Where the paper says it

• Performance and efficiency: Deconver achieves stateof-the-art performance on both 2D and 3D segmentation tasks with substantially fewer computational costs and parameters than leading baselines. Extensive experiments across four datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate Deconver’s superiority in Dice scores and boundary accuracy (Hausdorff distance).

Deconver achieves state-of-the-art performance on both 3D (ISLES’22 and BraTS’23) and 2D tasks with substantially fewer computational costs. p. 1
Where the paper says it

• Performance and efficiency: Deconver achieves stateof-the-art performance on both 2D and 3D segmentation tasks with substantially fewer computational costs and parameters than leading baselines. Extensive experiments across four datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate Deconver’s superiority in Dice scores and boundary accuracy (Hausdorff distance).

Performance and Efficiency

Deconver reduces FLOPs by up to 90% compared to attention-based baselines. For ISLES’22, Deconver (the variant using 3 × 3 × 3 kernel) achieves the highest DSC (78. p. 6

Where the paper says it

3: Comparison of Dice against the number of parameters (left) and FLOPs/voxel (right) for different models on ISLES’22. Deconver (the variant using 3 × 3 × 3 kernel) manages to maintain the highest DSC with fewer parameters and FLOPs/voxel.

16%) with fewer parameters and FLOPs/voxel. p. 5
Where the paper says it

Deconver requires over 70% fewer FLOPs per voxel compared to the SegResNet, the best performing baseline. Additionally, Deconver uses around 85% fewer parameters, resulting in a highly compact architecture without sacrificing segmentation performance.

In BraTS’23, Deconver variants outperform all baselines in average DSC across tumor subregions. p. 6
Where the paper says it

2) Brain Tumor Segmentation (BraTS’23): Table II presents segmentation performance on BraTS’23, comparing different models across three tumor subregions: enhancing tumor (ET), tumor core (TC), and whole tumor (WT). Both Deconver variants outperform all baselines in average DSC.

Ablation studies show that the best results are achieved when G is set to the number of channels and R = 4. p. 9
Where the paper says it

In general, these studies show that higher channel ratios improve segmentation quality but increase computational cost, while a higher number of groups significantly reduces the number of parameters while improving the performance. The best results are achieved when G is set to the number of channels and R = 4, as this configuration produces optimal segmentation accuracy with minimal overhead.

This configuration produces optimal segmentation accuracy with minimal overhead. p. 9
Where the paper says it

In general, these studies show that higher channel ratios improve segmentation quality but increase computational cost, while a higher number of groups significantly reduces the number of parameters while improving the performance. The best results are achieved when G is set to the number of channels and R = 4, as this configuration produces optimal segmentation accuracy with minimal overhead.

Conclusion

In this work, we introduce Deconver, a powerful segmentation network that integrates nonnegative deconvolution (NDC) as a learnable module within a U-shaped architecture. By replacing computationally expensive attention mechanisms with efficient deconvolution operations, Deconver restores high-frequency details while effectively suppressing artifacts. p. 9

Where the paper says it

CONCLUSION In this work, we introduce Deconver, a powerful segmentation network that integrates nonnegative deconvolution (NDC) as a learnable module within a U-shaped architecture. By replacing computationally expensive attention mechanisms with efficient deconvolution operations, Deconver restores highfrequency details while effectively suppressing artifacts.

Extensive experiments on four diverse medical imaging datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate that Deconver consistently achieves state-of-the-art segmentation performance. p. 9
Where the paper says it

By replacing computationally expensive attention mechanisms with efficient deconvolution operations, Deconver restores highfrequency details while effectively suppressing artifacts. Extensive experiments on four diverse medical imaging datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate that Deconver consistently achieves state-of-the-art segmentation performance, outperforming or matching leading CNNand Transformer-based models while significantly reducing computational costs.

Notably, Deconver reduces FLOPs by up to 90% compared to attention-based baselines. p. 9
Where the paper says it

Extensive experiments on four diverse medical imaging datasets (ISLES’22, BraTS’23, GlaS, and FIVES) demonstrate that Deconver consistently achieves state-of-the-art segmentation performance, outperforming or matching leading CNNand Transformer-based models while significantly reducing computational costs. Notably, Deconver reduces FLOPs by up to 90% compared to attention-based baselines, making it wellsuited for resource-constrained clinical applications.

Ablation studies further highlight the importance of key design choices, such as the source channel ratio and grouping strategy in NDC layers, in balancing accuracy and efficiency. p. 9
Where the paper says it

Notably, Deconver reduces FLOPs by up to 90% compared to attention-based baselines, making it wellsuited for resource-constrained clinical applications. Ablation studies further highlight the importance of key design choices, such as the source channel ratio and grouping strategy in NDC layers, in balancing accuracy and efficiency.

Deconver represents a promising step toward high-precision, computationally efficient medical image segmentation. p. 9
Where the paper says it

Ablation studies further highlight the importance of key design choices, such as the source channel ratio and grouping strategy in NDC layers, in balancing accuracy and efficiency. We believe Deconver represents a promising step toward high-precision, computationally efficient medical image segmentation, bridging the gap between classical image restoration and modern deep learning.

REFERENCE

Deconver introduces a novel network that integrates traditional deconvolution techniques from image restoration as a core learnable component within a U-shaped architecture. Deconver achieves state-of-the-art performance on both 2D and 3D segmentation tasks with substantially fewer computational costs and parameters than leading baselines. By bridging traditional image restoration with modern deep learning, Deconver provides a practical solution for high-precision segmentation in resource-constrained clinical workflows.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Nonnegative deconvolution layer allows for stable end-to-end training because it is backpropagation-friendly, differentiable layer based on a provably monotonic update rule for nonnegative deconvolution. This enables stable training using current deep learning frameworks.

  2. Deconver replaces computationally expensive attention mechanisms with efficient nonnegative deconvolution (NDC) operations, enabling the restoration of high-frequency details while suppressing artifacts. This allows the network to bridge traditional image restoration with deep learning.

  3. The multiplicative update rule for NDC is guaranteed to reduce the reconstruction error, as proven by Theorem 1: "the reconstruction error e(t) is non-increasing, i.e., e(t+1) <= e(t) for all t ≥ 0." This ensures stable and accurate restoration of high-frequency details.

  4. Deconver achieves up to 90% reduction in computational costs (FLOPs) compared to leading baselines, making it well-suited for resource-constrained clinical applications. This allows for practical deployment in environments with limited computational power.

  5. The architecture's design balances efficiency and context by replacing attention with a learnable Deconv Mixer module instead of attention, which is known to be computationally intensive, especially on long sequences. This maintains high performance while improving scalability.

  6. Ablation studies show that setting the number of groups G equal to the number of channels achieves the best DSC and HD95 while keeping the model compact. This indicates a strategy for maximizing segmentation accuracy with minimal overhead by optimizing group structure.

Abstract

While convolutional neural networks (CNNs) and vision transformers (ViTs) have advanced medical image segmentation, they face inherent limitations such as local receptive fields in CNNs and high computational complexity in ViTs. This paper introduces Deconver, a novel network that integrates traditional deconvolution techniques from image restoration as a core learnable component within a U-shaped architecture. Deconver replaces computationally expensive attention mechanisms with efficient nonnegative deconvolution (NDC) operations, enabling the restoration of high-frequency details while suppressing artifacts. Key innovations include a backpropagation-friendly NDC layer based on a provably monotonic update rule and a parameter-efficient design. Evaluated across four datasets (ISLES'22, BraTS'23, GlaS, FIVES) covering both 2D and 3D segmentation tasks, Deconver achieves state-of-the-art performance in Dice scores and Hausdorff distance while reducing computational costs (FLOPs) by up to 90% compared to leading baselines. By bridging traditional image restoration with deep learning, this work offers a practical solution for high-precision segmentation in resource-constrained clinical workflows. The project is available at https://github.com/pashtari/deconver.

Sources

Related papers