LW-MHFI-Net: a lightweight multi-scale network for medical image segmentation based on hierarchical feature incorporation

Medical image segmentation is an essential component in computer-aided diagnosis systems. It is about distinguishing the boundaries of targeted organs from close body tissue in CT or MRI images. Most medical images have helpful and critical information about targeted organs, like shape and size, which assists clinicians in diagnosing and treating different diseases. The encoder–decoder-based architectures are a powerful family of deep learning models and serve as spine models in different segmentation tasks. However, these prementioned architectures have two main drawbacks. First, there is diversity loss in the learned features because of using fixed-size filters and pooling operations in the utilized convolution kernel. Second, the requirement for a vast optimized network expands the number of parameters and accordingly the computational cost. Hence, this paper addresses these issues by proposing a lightweight multi-scale network, LW-MHFI-Net, to overcome the main demerits in the encoder–decoder-based structures. We propose an efficient multi-scale block called Hierarchical Feature Incorporation. This block is applied in a cascaded manner with dense connections at the bottleneck of the employed encoder–decoder-based network to enlarge the resultant receptive fields. The proposed model has fewer parameters, only 9.3 M, compared to the basic U-Net architecture, which has 31 M. Three public benchmarked datasets have been used to evaluate the performance of LW-MHFI-Net, including lung, skin lesions, and brain tumors datasets. The experimental results have shown that the proposed model outperforms the state-of-the-art methods in different metrics. Hence, the proposed LW-MHFI-Net can be applied as a robust tool in various medical image segmentation tasks.