This is the official implementation for article "LANet: Lightweight Attention Network for Medical Image Segmentation".
31.07.2024 - Sign the agreement, the article will be published on Springer
24.04.2024 - Attended the conference, waiting for final publication
11.03.2024 - The article is accepted and will be published after the conference which will be held on Azerbaijan.
20.12.2023 - The article is submitted in Springer proceedings of the ITTA-2024 conference (https://itta.cyber.az).
LANet, a Lightweight Attention Network, are presented in the paper and incorporates an Efficient Fusion Attention (EFA) block and an Adaptive Feature Fusion (AFF) decoding block. The model adopts MobileViT as a lightweight backbone network with a small number of parameters, facilitating easy training and faster predictive inference.
The EFA block enhances the model's feature extraction capability by capturing task-relevant information while reducing redundancy in channel and spatial locations.
The AFF decoding block fuses the purified low-level features from the encoder with the sampled features from the decoder, enhancing the network's understanding and expression of input features.
- torch == 2.1.1+cu121
- tensorboard == 2.11.2
- numpy == 1.24.1
- python == 3.9.18
- torchvision == 0.16.1+cu121
- ...
The efficiency of LANet was evaluated using four public datasets: kvasir-SEG, CVC-clinicDB, CVC-colonDB, and the Data Science Bowl 2018. All datasets used in paper are public, you can download online.
Split the datasets for train, validation and test with ratio 8:1:1
| Dataset | mDC | mIoU | mRec | mPrec |
|---|---|---|---|---|
| Kvasir-SEG | 0.911 | 0.851 | 0.903 | 0.949 |
| CVC_clinicDB | 0.944 | 0.896 | 0.926 | 0.966 |
| CVC_ColonDB | 0.771 | 0.712 | 0.758 | 0.894 |
| 2018 DSB | 0.930 | 0.871 | 0.918 | 0.946 |
waiting...
❗ 👀 The codes can not be used for commercial purposes!!!