Focus-UNet: Deep-stage Decomposition and Context Gating for Medical Image Segmentation

Abstract

Accurate segmentation of tumors and lesions in brain magnetic resonance imaging (MRI) is of great importance for clinical diagnosis, surgical planning, and prognosis assessment. However, brain lesions often exhibit substantial variations in size, irregular shapes, and blurred boundaries. Traditional U-Net relies on a bottom-up pyramidal feature extraction scheme and lacks explicit global semantic guidance, which may lead to false-positive predictions and discontinuous lesion boundaries. To address these limitations, this paper proposes Focus-UNet, a brain MRI segmentation network based on an "overview-first, focus-next" mechanism. The network decomposes the conventional U-Net encoder into three components: Base-Net, Overview-Net, and Focus-Net. Base-Net extracts shallow-and mid-level features; Overview-Net rapidly compresses the feature maps through large-stride convolutions to generate a global context prior containing overall anatomical information and approximate lesion locations; and Focus-Net adaptively modulates local convolutional features under the guidance of this prior through a dynamic context gating mechanism, thereby enhancing lesion-related responses while suppressing irrelevant background noise. The decoder progressively restores spatial resolution through upsampling and multi-scale skip connections, integrating deep semantic features from Focus-Net with high-resolution detailed features from Base-Net to recover lesion structures and boundaries. Experimental results on the Brain MRI dataset demonstrate that the proposed Focus-UNet achieves better segmentation performance than the original U-Net, confirming the effectiveness of the global overview prior and dynamic context gating mechanism for brain MRI lesion segmentation.