Medical Image Recognition Based On Deep Learning

Pulmonary nodule is a multi-systems and multi-organs disease and may be a symptom of early lung cancer.Global lung cancer's incidence and mortality rates is 13% and 19% of all cancers.Due to the high fatality rate of lung cancer,the diagnosis of pulmonary nodules is particularly important.Traditional method of pulmonary nodules diagnosis uses no aided eyes of human being to look at the lung CT slices and recognize.But a patient usually has hundreds of CT slices and some pulmonary nodules may be very small.So Computer-aided diagnosis can help doctors to recognize the nodules faster and can save the precious time and energy of them.A method of pulmonary nodules recognition based on deep learning is proposed in this paper.We combined deep learning with medical images and use deep learning to analyze preprocessed image and extract some features that may not be found by human eyes automatically.In the phase of pulmonary detection,since the lack of 2D CT images' feature,a method of image preprocessing based on RGB channels superposition is proposed to enhance the features.The image preprocessing steps are as follows.First,we extract the lung parenchyma to from raw CT images to reduce the effect of lung outline to the experiment results.Then we extract the regions of interest.In pulmonary nodules experiments,we use the method of ROI superposition to get pseudo-color images which can show the trend of tissues in vertical in order to enhance the difference between healthy tissues and pulmonary nodules,especially for some small nodules.After preprocessing,we use deep learning methods to train these data and obtain a prediction model to test.The experiments use data in LIDC-IDRI database and GGO nodules data provided by XINHUA hospital.The pulmonary nodules detection experiment gains the sensitive of 95.0% at average 5.62 false positives per scan and the GGO nodules classification experiment gets the best F-score of 0.87805.