基于APC-UNet模型的皮肤病变区域分割研究

Research on Skin Lesion Region Segmentation Based on APC-UNet Model

针对皮损皮肤镜图像分割不准确的问题,本文提出了一种基于Atrous-spatial-pyramid-pooling Parallel Coordinate-attention pattern U-Net(APC-UNet)模型的皮肤病变区域分割算法。算法在U-Net模型的编码器中融入Atrous Spatial Pyramid Pooling(ASPP)模块和ParNet模块以提升模型的特征提取能力,在解码器中嵌入带有注意力机制的Coordinate Attention(CA)模块以增强模型的定位能力,并且引入了Lovász-hinge损失函数来解决皮损皮肤镜图像样本类别不均衡的问题。通过消融实验验证了提出的模型的改进合理性,通过对比实验结果表明,APC-UNet模型整体上优于5种对比模型,并且相较于基准模型U-Net,在Dice系数、IoU、精确率、召回率和准确度上分别提升了6.14%、8.11%、6.79%、2.28%和2.49%,各项性能指标均有较好提升,是一种有效可行的皮肤病变区域分割算法。

Aiming at the problem of inaccurate segmentation of skin lesion dermatoscope images,a skin lesion region segmentation algorithm based on Atrous-spatial-pyramid-pooling Parallel Coordination pattern U-Net(APC-UNet)model has been proposed in the present study.The Atmosphere Spatial Pyramid Pooling(ASPP)modules and ParNet modules are integrated into the encoder of the U-Net model in the algorithm to improve the feature extraction capability of the model.And the Coordinate Attention(CA)modules are embeded with attention mechanism in the decoder to enhance the positioning capability of the model.In addition,the Lovász-hinge loss function is introduced to solve the problem of class imbalance of skin lesion dermatoscope images.The improvement rationality of the proposed model through the ablation experiment is verified.The comparison experiment results show that the APC-UNet model is better than the five comparison models on the whole,and compared with the benchmark model of U-Net,the Dice coefficient,IoU,precision,recall and accuracy are improved by 6.14%,8.11%,6.79%,2.28%and 2.49%,respectively,and all performance indicators are improved,which is an effective and feasible skin lesion region segmentation algorithm.