Official PyTorch implementation of our paper HyenaMed: Lightweight Skin Lesion Segmentation via Convolutional Global Context Modeling, presented at the 2026 International Conference on Frontiers of Engineering and Emerging Technologies (FET), Sakhir, Bahrain.
Z. Haider, H. Haiyu, M. A. F. Butt and M. Ali, "HyenaMed: Lightweight Skin Lesion Segmentation via Convolutional Global Context Modeling," 2026 International Conference on Frontiers of Engineering and Emerging Technologies (FET), Sakhir, Bahrain, 2026, pp. 1-7, doi: 10.1109/FET68771.2026.11601590.
Accurate skin lesion segmentation is a critical step toward automated dermoscopic diagnosis, but most high-performing segmentation models rely on heavy backbones that are impractical for deployment on resource-constrained or point-of-care devices. HyenaMed addresses this gap by adapting convolutional global-context modeling (inspired by HyenaPixel) into an extremely lightweight encoder-decoder architecture for skin lesion segmentation, achieving competitive accuracy at a fraction of the computational cost of standard CNN and transformer-based baselines.
Key highlights:
- ~0.68M parameters and ~0.30 GFLOPs — orders of magnitude smaller than typical segmentation backbones
- Large effective receptive field via convolutional global context modeling, without the quadratic cost of self-attention
- Evaluated on three public dermoscopic benchmarks: ISIC 2016, ISIC 2017, and ISIC 2018
- Built on top of MMSegmentation for reproducible training, testing, and benchmarking
HyenaMed is a lightweight encoder-decoder segmentation network. The encoder stacks four HPx-Blocks at progressively downsampled resolutions (H/4, H/8, H/16, H/32), each combining a LayerNorm → HPx (a convolutional global-context mixer inspired by Hyena, using depth-wise/point-wise convolutions with an implicit positional filter and long-range global convolution) → LayerNorm → FFN, with 3×3 stride-2 convolutions merging features between stages. Multi-scale features are projected to a common channel dimension by a shared MLP layer and upsampled to a common resolution. The decoder fuses these features through a PBRP+ADWP module: a Point-wise Bottleneck Residual Path (PW Conv 1×1 → BatchNorm → ReLU → PW Conv 1×1) combined with an Attentive Dual-Window Pooling branch (parallel n×1 and 1×n depth-wise convolutions, followed by average/max pooling gated through a sigmoid attention weight), before a final MLP head predicts the segmentation mask.
| Dataset | aAcc | mIoU | mDice | mPrecision | mRecall | Params | GFLOPs |
|---|---|---|---|---|---|---|---|
| ISIC2016 | 96.67 | 89.51 | 90.34 | 94.54 | 94.15 | 0.68M | 0.30G |
| ISIC2017 | 92.65 | 83.63 | 90.95 | 91.93 | 90.10 | 0.68M | 0.30G |
| ISIC2018 | 91.85 | 82.56 | 90.31 | 89.95 | 90.70 | 0.68M | 0.30G |
Full comparisons against DeepLabV3+, MobileNetV3, Fast-SCNN, and SegFormer baselines are provided in the paper.
conda create -n hyenapixel python=3.10
conda activate hyenapixel
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -e .Download the ISIC 2016, ISIC 2017, and ISIC 2018 lesion segmentation datasets from the ISIC Archive and arrange them in MMSegmentation format:
segmentation/data/
├── isic2016/
│ ├── images/
│ │ ├── training/
│ │ └── validation/
│ └── annotations/
│ ├── training/
│ └── validation/
├── isic2017/
│ └── ...
└── isic2018/
└── ...
Segmentation masks should use the suffix _Segmentation.png, matching the dataset config files under mmsegmentation/configs/_base_/datasets/.
All commands are run from the segmentation/ directory.
python mmsegmentation/tools/train.py mmsegmentation/configs/hpx_former_s18/segformer_hpx-former-s18-90k_isic2016-224x224.py
python mmsegmentation/tools/train.py mmsegmentation/configs/hpx_former_s18/segformer_hpx-former-s18-90k_isic2017-224x224.py
python mmsegmentation/tools/train.py mmsegmentation/configs/hpx_former_s18/segformer_hpx-former-s18-90k_isic2018-224x224.pyMulti-GPU training:
CUDA_VISIBLE_DEVICES=0,1 bash mmsegmentation/tools/dist_train.sh mmsegmentation/configs/hpx_former_s18/segformer_hpx-former-s18-90k_isic2016-256x256.py 2python mmsegmentation/tools/test.py \
mmsegmentation/configs/hpx_former_s18/segformer_hpx-former-s18-90k_isic2016-224x224.py \
work_dirs/segformer_hpx-former-s18-90k_isic2016-224x224/iter_90000.pth \
--show-dir results/segformer_hpx-former-s18-90k_isic2016-224x224/test/ \
--work-dir results/segformer_hpx-former-s18-90k_isic2016-224x224/test/python mmsegmentation/tools/analysis_tools/get_flops.py \
mmsegmentation/configs/hpx_former_s18/segformer_hpx-former-s18-90k_isic2016-224x224.py --shape 256python mmsegmentation/tools/analysis_tools/benchmark.py \
mmsegmentation/configs/hpx_former_s18/segformer_hpx-former-s18-90k_isic2016-224x224.py \
work_dirs/segformer_hpx-former-s18-90k_isic2016-224x224/iter_90000.pthWe benchmark HyenaMed against the following baselines, trained under the same protocol using MMSegmentation:
- DeepLabV3+
- MobileNetV3
- Fast-SCNN
- SegFormer
Configuration files for each baseline are available under mmsegmentation/configs/.
Our implementation builds on and is grateful to the following open-source projects:
- HyenaPixel: Global Image Context with Convolutions
- MambaU-Lite: A Lightweight Model based on Mamba and Integrated Channel-Spatial Attention for Skin Lesion Segmentation
- MMSegmentation
If you find this work useful for your research, please cite:
@inproceedings{haider2026hyenamed,
author = {Z. Haider and H. Haiyu and M. A. F. Butt and M. Ali},
title = {HyenaMed: Lightweight Skin Lesion Segmentation via Convolutional Global Context Modeling},
booktitle = {2026 International Conference on Frontiers of Engineering and Emerging Technologies (FET)},
address = {Sakhir, Bahrain},
pages = {1--7},
year = {2026},
doi = {10.1109/FET68771.2026.11601590}
}This project is released under the MIT License.