Deconver: A Deconvolutional Network for Medical Image Segmentation

TL;DR

Deconver tackles the trade-off between local CNNs and global but computation-heavy Transformers by embedding a nonnegative deconvolution (NDC) layer as a core learnable module within a U-Net–style architecture. It replaces attention with a Deconv Mixer that leverages NDC to restore high-frequency details while suppressing artifacts, accompanied by a provably monotonic multiplicative update and a compact, group-wise design. Across four datasets (ISLES'22, BraTS'23, GlaS, FIVES) for 2D and 3D segmentation, Deconver achieves state-of-the-art Dice scores and competitive boundary metrics while reducing FLOPs by up to 90% and shrinking parameter counts substantially relative to baselines. The approach bridges traditional image restoration with deep learning to deliver high-precision segmentation suitable for resource-limited clinical workflows, and it provides a practical framework for integrating deconvolution concepts into modern segmentation models.

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.

Figures (6)

• Figure 1: Overview of Deconver architecture.
• Figure 2: Overview of Deconver block and its components.
• Figure 3: Comparison of Dice against the number of parameters (left) and FLOPs/voxel (right) for different models on ISLES'22. Deconver (the variant using $3 \times 3 \times 3$ kernel) manages to maintain the highest DSC with fewer parameters and FLOPs/voxel.
• Figure 4: Qualitative results of stroke lesion segmentation on ISLES'22. The regions of true positives are marked in blue, false positives in green, and false negatives in red. The DSC score is presented for each case. SegResNet introduces false positives in the center of the image in the first row (marked with the orange circle), while Swin UNETR, UNETR, and nnU-Net consistently under-segment the lesion in both examples.
• Figure 5: Qualitative results of brain tumor segmentation on BraTS'23. Tumor core (TC) is the union of red (NCR/NET) and yellow (ET) regions, and whole tumor (WT) is the union of green (edema), red, and yellow regions. Each row displays a sample slice from a subject in the validation set. The average DSC score is presented for each case. In the first row example, all of the baselines fail to detect part of the enhancing tumor marked by the orange circle. In the second row example, nnU-Net, Swin UNETR and UNETR do not capture fully the edema, while SegResNet and Swin UNETR falsy predict the area marked by the orange circle as edema.

Theorems & Definitions (2)

• Theorem 1: Monotonicity
• proof