MFDC-Net:一种融合多尺度特征和注意力机制的乳腺癌病理图像分类算法

MFDC-Net:A breast cancer pathological image classification algorithm incorporating multi-scale feature fusion and attention mechanism

乳腺癌是全球最常见的恶性肿瘤之一,采用传统方法诊断需花费大量时间和精力,且受个人能力影响较大。用计算机辅助诊断的方法,可以提高病理图像分类的准确率和效率,从而满足临床应用的需求。为此,提出一种基于DenseNet的融合多尺度特征和注意力机制的乳腺癌病理图像分类算法(MFDC-Net)。在密集块中引入坐标注意力机制,精准定位重要特征的空间信息。采用多尺度池化过渡层,通过不同卷积核的平均池化和普通卷积,在实现降维的同时扩大感受野。采用多尺度特征增强模块,融合深层次图像特征,提高分类性能。结果显示,MFDC-Net模型的分类性能较其他经典模型更优,分类准确率达97.12%,易混淆率低至3.34%,能较好地进行乳腺癌组织病理图像分类,为诊断和治疗提供重要依据。

Breast cancer is one of the most common malignant tumors in the world.Traditional methods take pathologists a lot of time and effort to diagnose,and the results are greatly affected by individual abilities.Using computer-aided diagnosis methods can improve the accuracy and efficiency of pathological image classification,meet the demands of clinical applications.To this end,a multi-scale feature fusion based on DenseNet and coordinate attention network(MFDC-Net)is proposed.The introduction of coordinate attention mechanism into the dense blocks can locate important feature spatial information precisely.The improved transition layers use average pooling and normal convolutions with different convolution kernels to reduce dimension and expand receptive fields.Finally the improved network employs a multi-scale feature fusion model using dilated convolution,average pooling and normal convolutions to fuse deep image features to improve classification performance.The experimental results show that MFDC-Net model has better classification performance,the accuracy rate of four classifications reaches 97.12%,the easily confused rate decreases to 3.34%.The method can better classify the histopathological images of breast cancer,and can provide an important basis for the diagnosis and treatment of doctors.