Striking a better balance between segmentation performance and computational costs with a minimalistic network design

While many modern segmentation models achieve appealing performance, they often come with high computational costs, such as massive parameter counts, intensive calculations, excessive memory consumption, and prolonged runtime. These burdens make them impractical for resource-limited medical devices. To address this challenge, we strategically integrate several established modules, including asymmetric convolution, dilated convolution, and attention modules, to develop a novel medical image segmentation backbone named BMIS. Specifically, unlike traditional U-Nets, BMIS adopts a shallower yet wider structure. This design not only enables effective learning of high-level semantic features but also markedly mitigates the loss of fine-grained details caused by frequent downsampling. Consequently, the decoder in BMIS can focus primarily on spatial reconstruction rather than compensating for lost details, which greatly reduces its workload. Leveraging these benefits, the decoder of BMIS does not need to be as complex as its encoder, and skip connections between the encoder and decoder are no longer necessary. These two factors collectively reduce model complexity without compromising segmentation performance. Extensive segmentation experiments across five medical imaging modalities demonstrate that BMIS achieves a better balance between segmentation performance and computational costs compared to twenty-nine competing methods. In short, compared with the best-performing comparative method DSU-Net, BMIS improves IoU by ∼0.32%, while substantially reducing model parameters by 97.54%, FLOPs by 68.49%, GPU memory usage by 54.48%, training time by 58.82%, and inference time by 80.00%. These impressive results highlight BMIS's great potential for deployment in medical devices with limited computational resources.