HALF-MAFUNET: A Lightweight Architecture Based on Multi-Scale Adaptive Fusion for Medical Image Segmentation

Maula Sandy, Abiaz Fazel
Prasetyo, Heri

Journal of Electronics, Electromedical Engineering, and Medical Informatics
Vol 8 No 1 (2026): January

Medical image segmentation is a critical component in computer-aided diagnosis systems but many deep learning models still require large numbers of parameters and heavy computation. Classical CNN-based architectures such as U-Net and its variants achieve good accuracy, but are often too heavy for real deployment. Meanwhile, modern Transformer-based or Mamba-based models capture long-range information but typically increase model complexity. Because of these limitations, there is still a need for a lightweight segmentation model that can provide a good balance between accuracy and efficiency across different types of medical images. This paper proposes Half-MAFUNet, a lightweight architecture based on multi-scale adaptive fusion and designed as a simplified version of MAFUNet. The main contribution of this work is combining the efficient encoder structure of Half-UNet with advanced fusion and attention mechanisms. Half-MAFUNet integrates Hierarchy Aware Mamba (HAM) for global feature modelling, Multi-Scale Adaptive Fusion (MAF) to combine global and local information, and two attention modules, Adaptive Channel Attention (ACA) and Adaptive Spatial Attention (ASA), to refine skip connections. In addition, this model incorporates Channel Atrous Spatial Pyramid Pooling (CASPP) to capture multi-scale receptive fields efficiently without increasing computational cost. Together, these components create a compact architecture that maintains strong representational power. The model is trained and evaluated on three public datasets: CVC-ClinicDB for colorectal polyp segmentation, BUSI for breast tumor segmentation, and ISIC-2018 for skin lesion segmentation. All images are resized to 256×256 pixels and processed using geometric and intensity-based augmentations. Half-MAFUNet achieves competitive performance, obtaining mean IoU around 84 85% and Dice/F1-Score around 90 92% across datasets, while using significantly fewer parameters and GFLOPs compared to U-Net, Att-UNet, UNeXt, MALUNet, LightM-UNet, VM-UNet, and UD-Mamba. These results show that Half-MAFUNet provides accurate and efficient medical image segmentation, making it suitable for real-world deployment on devices with limited computational resources.