多深度和多尺度特征融合的乳腺癌病理图像分类OA

Breast cancer pathological images contain complex visual features such as extensive cell overlap and interlacing,with subtle differences among images.Extracting these features using a single network structure is challenging,and simply deepening the network may lead to gradient vanishing.This study adopts a multi-network feature fusion approach,combining the efficient feature reuse capability of DenseNet with the multi-scale convolutional advantages of the Inception architecture.This integration enables effective extraction and fusion of features across different scales and depths.Subsequently,decision-level feature fusion is further performed on the features extracted by both networks.This method can extract rich features more effectively and make fusion decisions,which improves the perception ability of the algorithm.Comparative experiments demonstrate that the accuracy rate of the proposed fusion network achieves 98.61%and its recall rate achieves 99.21%in breast cancer pathological image classification.The proposed network outperforms single-network feature extraction models such as FE-BkCapsNet and AlexNet,as well as multi-network feature fusion models like ResHist and AlexNet+VGG16,exhibiting superior classification performance.Furthermore,ablation experiments indicate that the accuracy rate,precision,recall rate,and F1-score of Inception+DenseNet fusion network are significantly improved in comparison with those of the single-network models such as Inception and DenseNet.This fusion network fully leverages the complementary advantages of Inception and DenseNet in feature extraction,and achieves a synergistic gain effect,which further validates its effectiveness and superiority in handling complex pathological images.