Research Of Medical Image Classification Based On Deep Learning

With the development of society,automobile exhaust and industrial air pollution are increasing day by day.Lung cancer has become one of the malignant tumors that threaten the health of people's lives.Chest imaging is a commonly used method for lung cancer detection.Through the patient's chest radiograph,bone scan,CT examination,MRI examination and other technical means,we can clearly observe the patients' lungs overall information and lesion information.The size and shape of the nodules in the lung medical images are important basis for detecting lung cancer.How to accurately locate the location of nodules from the pictures of multiple lungs of patients and evaluate the degree of nodule deterioration are the key technologies of computer-aided diagnosis of lung cancer.Traditional deep learning and detection of lung medical images often use two-dimensional convolution kernel,ignore the single patient's lung image in the two frames before and after the spatially related information.The disadvantages result in incomplete extraction of lung node features,the loss of accuracy.Aiming at the problems and defects of deep learning detection algorithm with two-dimensional convolution kernel,this paper proposes a new classification algorithm of lung cancer node detection based on three-dimensional convolution kernel,in which the network input composed of a single patient's lung image changes to multiple images input at the same time.According to the relative steps of moving in different directions,taking the size of the fixed segmentation window,shearing the input multiple images,converting multiple images input by one patient into multiple fixed-window-sized three-dimensional convolution images as a network image input,and calculating the results according to depth learning network weights,we determine whether there are nodule nodes and the position by the input three-dimensional images.Since the size of splitting 3D image window is fixed,the size of the input image is adjustable,and multiple images are inputted at the same time,the operation time of sequentially executing multiple images may be reduced and the speed may be accelerated.In this paper,a large number of public lung cancer nodules datasets were used to train the network offline.The training set is consisted of a total of 858 lung CT images of 885 patients,including 1180 lung cancer node positions.Using a large number of lung cancer patients' offline images to train neural network,the network has a strong generalization ability,and the test set has a good calculation accuracy.The algorithm accelerates the computation with GPU,which can accelerate the speed and accuracy very well,can control the time of processing hundreds of lung images in a single patient within 60 seconds and maintain a good calculation accuracy.In this paper,we compare the proposed algorithm with the other four different neural network algorithms.The results show that the proposed algorithm has a higher classification accuracy and more ideal processing speed.