Breast pathological image diagnosis algorithm incorporating adaptive feature fusion and conditional random field
Pathological analysis is one of the common methods for cancer diagnosis.Although pathological examination based on deep learning exhibits good performance,the processing method for tissue slices tends to ignore the spatial correlation of pathological tissues.In order to obtain breast cancer classification results and malignant tumor location more accurately,a Transformer framework embedded with adaptive feature fusion module and mean value conditional random field is proposed,and the whole framework is trained end-to-end using back propagation algorithm.The adaptive feature fusion module uses learnable parameters to combine the improved self-attention and multi receptive field convolution module adaptively for obtaining multi-scale semantic features and enhancing the model feature extraction capability from both global and local perspectives.The proposed mean value conditional random field is combined with the backbone network to integrate the spatial correlation between tissue slices and obtain morphological information between pathological tissues.Experimental results show that the proposed method yields 95.51%accuracy on slice images,and achieves 0.9745 AUC and 0.8102 FROC on whole-slice images,demonstrating its feasibility and higher diagnostic accuracy for pathological image classification.
breastimage processingadaptive feature fusionconditional random fieldpathological slice
NETL
NSTL
万方数据
维普
