An Attention-Enhanced U-Net Model for Histopathology Image Segmentation

Abstract

Accurate image segmentation remains a challenging task, particularly in complex domains such as medical imaging, where precise boundary detection and feature discrimination are critical. In this work, an attention-enhanced U-Net architecture is proposed to improve segmentation performance by integrating multiple attention mechanisms in the encoder as well as in the decoder framework. The encoder incorporates the Convolutional Block Attention Module (CBAM) to strengthen feature extraction by focusing on informative spatial regions and channels. The bottleneck layer captures high-level semantic representations with increased feature depth, followed by dropout regularization to reduce overfitting. In the decoder, Squeezeand-Excitation (SE) blocks are applied to recalibrate channelwise feature responses during reconstruction while preserving fine spatial details through skip connections. The proposed model produces pixel-wise segmentation maps using a convolution layer and is trained employing a binary cross-entropy based loss. The model is tested on two standard datasets, TNBC and MoNuSeg, and has achieved Dice scores of $\mathbf{8 4. 3 0} \boldsymbol{\pm} \mathbf{0 0. 5 0} \boldsymbol{\%}$ and $\mathbf{8 1. 0 6} \boldsymbol{\pm} \mathbf{0 0. 3 0} \boldsymbol{\%}$, respectively, and IoU of $\mathbf{7 3. 4 5} \boldsymbol{\pm} \mathbf{0 0. 4 0} \boldsymbol{\%}$ and $\mathbf{6 8. 2 3} \boldsymbol{\pm} \mathbf{0 0. 4 0} \boldsymbol{\%}$, respectively. Experimental results validate that the fusion of attention mechanisms enhances feature representation and improves overall segmentation accuracy compared to existing methods.