Understanding Medical Image Based On Structured Learning

In clinical practice,professional doctors generally evaluate medical images through subjective observations.This assessment relies on experience accumulation and there are individual differences.Compared to this qualitative reasoning,artificial intelligence is good at identifying complex patterns in data and providing quantitative assessments in an automated way.It is very meaningful to introduce image understanding into the medical field to generating semantic annotation and description for medical images.In this paper,we focus on mammography processing,and the method applied to image understanding is based on structured learning.Through the research on the detection of lesion area,label generation and semantic mapping,we implemented the structured output of mammography.The main work of this paper is:(1)Due to the difficulty of obtaining manual labeling,an automatic lesion detection algorithm based on unsupervised learning was proposed.The algorithm designed with K-means clustering base method,extracted pixel feature and classify them.After the lesions were initially clustered,the morphological and texture features of the regions are extracted.The support vector machine(SVM)was used for post-processing filter,and finally the region of interest(ROI)was got.(2)In order to reduce the number of proposal regions in the multi-label classification(MLC)task and improve the feature extraction ability,a novel method based on ROI Crop Pooling(RCP)was proposed for feature extraction network.It can more accurately identify the characteristics of medical images with few model parameters.(3)Based on the previous two steps,for resolve the over-fitting problem that is easily caused by small-sample training,this paper applied a Full Convolutional Network(FCN)in image description generation framework,called FCN-MLC-LSTM,which is proposed to transfer image visual feature to natural statement description.This paper also introduces post-processing operations such as beam search to make the statements generated by the model closer to the natural representation.This paper is based on the open dataset and the breast clinical data of a top three hospital in Shanghai.The experimental results show that the proposed methods have achieved good results on all three tasks.