Research Of Weakly Supervised Fine-grained Part-based Image Classification Algorithm

Breast tumor is a major disease that endangers women’s health,and early ultrasound diagnosis of breast tumors is an important method to prevent and diagnose tumors.Clinically,doctors mainly make preliminary diagnosis based on the BI-RADS classification results of breast tumors.BI-RADS classification of breast ultrasound images is different from benign and malignant coarse-grained classification tasks.It is a fine-grained classification problem.It faces multiple challenges such as large inter-class variance,small intra-class variance and large morphological effects.The collection and labeling of medical image data sets requires professional physicians,and the workload is extremely high and costly.In order to reduce physician labeling pressure and get closer to actual engineering conditions,this article mainly discusses the fine-grained BI-RADS classification of breast ultrasound images under weakly supervised settings.The resolution of fine-grained image classification problems can be roughly divided into three directions: fine-tuning methods based on conventional image classification neuron networks,methods based on fine-grained feature learning,and methods based on object parts.In this paper,we discuss part-based methods.There are two main technical points in fine-grained image classification: one is to filter out the informative areas from the picture,and then convert the informative areas into discriminative areas.The discriminative areas are more conducive to learning differences between categories than the informative areas,which better helps fine-grained classification;the second is to combine the characteristics of the dataset with the characteristics of the classification problem,and introduce or mine prior information to help the classifier learn.Fully combining the characteristics of the data set and the characteristics of the classification problem can often achieve a multiplier effect.According to the technical point of introducing or mining prior information from the dataset,we elaborated and introduced two light and portable methods: partial cross entropy and soft labeling.In the direction of transforming informative regions into discriminative regions,we propose a novel DEL module,and by visualization we show that our proposed method can make the network pay attention to the discriminative regions.We further propose an interpretable fine-grained breast BI-RADS classification algorithm based on the former two ways.By extracting features according to different pathological parts of target tumor,we can obtain interpretable pathological features.At the same time,we can output the visualized results of the rough location of the tumor and the pathological location corresponding to the extracted features,and obtain a more reasonable pathological interpretation by visualizing the different locations of the tumor.Our visualization results show that the proposed algorithm can roughly find the target location of the tumor,and the proposed part-based pathological features roughly correspond to accurate locations,which can form a corresponding relationship with the actual basis for doctors to make diagnostic judgments.