Research On Classification Of Breast Histopathological Images Based On Feature Fusion

Large-scale training data and stacked layers make convolutional network models more efficient for classifying natural images.Compared with natural images,the complexity and diversity of breast tissue pathological images have brought a lot of obstacles to the improvement of classification accuracy.How to use the feature information of pathological image to provide reliable basis for intelligent diagnosis has become the focus of many scholars.At present,most of the classification models only classify a single image,without considering the distribution of multiple pathological images from the same patient in the medical dataset.However,intelligent diagnosis ultimately needs to be implemented to classify each patient’s disease.If the pathological image classification results of the same patient under different magnification are different,it will inevitably affect the final diagnosis results.Therefore,it is of great significance to study the multi-channel feature fusion of pathological images.This paper is oriented to intelligent diagnosis,and researches the classification of breast tissue pathological dataset from the perspective of images and patients,a multichannel fusion classification method combining dense connection and spatial attention mechanism is proposed.The main contents of this paper are as follows:(1)Preprocessing of breast tissue pathology images and comparison of classification methods.Aiming at the imbalanced distribution of breast tissue pathological dataset caused by the difficulty of patient image collection,a data preprocessing method based on the morphological structure and texture feature distribution of pathological images is proposed,which can enhance the data according to the distribution of each category in the original dataset.Based on this,considering the influence of sparse chromatin and blurred texture features on feature extraction due to improper operation during pathological section staining,a pathological image with standard staining is selected as a reference,and other images are staining normalized in color space.Besides,the traditional image classification method and image classification method based on deep learning are compared to select the more suitable image classification method.(2)Classification of breast histopathological images by combining multidimensional features with spatial attention mechanism.In order to solve the problem of insufficient feature reuse and weak spatial adaptability in the process of image classification,combined with transfer learning,the dense connection and spatial attention mechanism are leaded into the pathological image classification method at the same time.To strengthen the reuse of the superficial and deep features of pathological images and improve the spatial adaptability of classification model,experiments are designed to compare the different effects of using different feature fusion strategies and embedding the spatial attention mechanism in different convolution layers.(3)Breast histopathological image classification based on the fusion of patient characteristics.Cascaded network model is trained respectively using pathological images with different magnification in the breast histopathological dataset,and the pathological images are reclassified according to the patient number.The corresponding cascade model is used to classify the pathological images from the same patient at different magnifications.Finally,an ensemble algorithm is designed to fuse the prediction results of the base model to make full use of the patient’s features at different magnifications,and to improve the accuracy of patient-oriented pathological image classification.