Reasearch On Feature Combination And Feature Learning For Medical Image Classification

In last decades,with the rapid development and popularization of medical imaging equipment,medical image data has increased explosively.It has become a great challenge to analyze medical image data efficiently and accurately.On the other hand,computer aided detection/diagnosis has been proposed to improve the efficiency and accuracy of medical imaging analysis,and becomes an increasingly hot research area.In recent years,deep learning technology has achieved great success in the fields of image recognition,target detection etc.,and also been applied to medical image analysis.The target of this work is to apply deep learning technology to automatic detection of lesions in medical images.Two applications are considered,one is the recognition of pulmonary nodules in chest radiograph,and the other is to identify CT abdominal lymph nodes.Due to a large number of parameters involved in a deep network,a large number of labeled data are needed during training.But,labeled medical images are always rare.This work uses two methods to solve this problem.The first is to use unsupervised feature learning to extracted features of lesions.Another is deep learning based transfer-learning,which trains a deep network on dataset from other domain,and fine-tune it in the target medical image domain.The main contributions of this paper are as follows:(1)an automatic algorithm is proposed for chest lung nodules recognition,which uses Sparse Convolution Encode for feature extraction.The new feature can achieve comparable accuracy to traditional hand-crafted features,and combining the two kinds of features can achieve better performance;(2)for the detection of CT abdominal lymph nodes,we train a stacked Convolution Auto-Encode algorithm model in a unsupervised fashion to extract lymph features,and then train the classifier with labeled data;(3)we also propose a lymph node identification algorithm based on deep transfer learning,which first trains a CNN on a large number of natural scene images(ImageNet),then fine-tune the model based on labeled CT lymph nodes data.