This repository is the official implementation of "Deconver: A Deconvolutional Network for Medical Image Segmentation".

---

Deconver is a segmentation architecture inspired by deconvolution techniques. We introduce a novel mixer module based on nonnegative deconvolution (NDC), which effectively restores high-frequency details while suppressing artifacts. Built on a U-shaped backbone, Deconver replaces computationally expensive self-attention blocks with this mixer to efficiently capture special dependencies.

✨ Key Features

• 🏆 State-of-the-art performance on various 2D/3D medical segmentation tasks (ISLES'22, BraTS'23, GlaS, FIVES)
• ⚡ Significantly fewer FLOPs compared to CNN and Transformer baselines
• 🧠 Parameter-efficient mixer module using a multiplicative update rule for the source in deconvolution

🔥 Latest News

• [April 2, 2024] The preprint of our paper is available on arXiv!
• [March 26, 2024] The source code for Deconver is released!

🛠️ Installation

Install Deconver directly from GitHub using:

pip install git+https://github.com/pashtari/deconver.git

🚀 Quick Start

2D Segmentation: GlaS Example

import torch
import torch.nn as nn
from deconver import Deconver

model = Deconver(
    in_channels=3,
    out_channels=1,
    spatial_dims=2,
    encoder_depth=(1, 1, 1, 1, 1),
    encoder_width=(32, 64, 128, 256, 512),
    strides=(1, 2, 2, 2, 2),
    decoder_depth=(1, 1, 1, 1),
    norm=nn.InstanceNorm2d,
    kernel_size=(5, 5),
    groups=-1,              # depth-wise grouping
    ratio=4,                # source channel expansion ratio
)

x = torch.rand(1, 3, 256, 256)
y = model(x)                # output logits

3D Segmentation: ISLES'22 Example

model = Deconver(
    in_channels=2,
    out_channels=1,
    spatial_dims=3,
    encoder_depth=(1, 1, 1, 1),
    encoder_width=(64, 128, 256, 512),
    strides=(1, 2, 2, 2),
    decoder_depth=(1, 1, 1),
    norm=nn.InstanceNorm3d,
    kernel_size=(3, 3, 3),
    groups=-1,
    ratio=4, 
)

x = torch.rand(1, 2, 64, 64, 64)
y = model(x)

📊 Results

Dataset	DSC (%) ↑	HD95 ↓	Params (M)	FLOPs / pixel (K)
ISLES'22	78.16	4.99	10.5	607.0
BraTS'23	90.66	4.45	10.6	167.5
GlaS	92.12	60.49	20.8	422.8
FIVES	92.72	30.26	20.8	422.8

Metrics: Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95)

📜 Citation

@article{ashtari2024deconver,
  title={Deconver: A Deconvolutional Network for Medical Image Segmentation},
  author={Ashtari, Pooya and others},
  journal={arXiv preprint arXiv:2504.00302},
  year={2024}
}

📄 License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

📬 Contact

For questions or collaboration, contact:

• Pooya Ashtari: pooya.ash@gmail.com
• GitHub: @pashtari