Research Of Lung Tissue Classification Based On Deep Learning

Interstitial Lung Disease (ILD) is a clinical umbrella term referring to a diverse range of lung disorders that involve inflammation and fibrosis of interstitium. The most common symptoms of ILD include labored breathing, dry cough and shortness of breath, which produce seriously adverse impact on our lives. Almost all patients will need a high-resolution computed tomography (HRCT) scan of the chest to visualize the subtle texture changes of different ILD lung tissues, in which computer vision tends to work better than clinical radiologists in detecting these differences. Due to the inherent complexity of interpretation of HRCT images, a computerized approach for differentiating the lung tissues to provide helpful suggestions would be appealing for lung specialists.In order to improve the performance of lung tissue classification system, in this thesis, we propose some feature learning methods. The main research contents are stated as follows:The mainly previous works focused on lung texture classification were based on hand-crafted features. They designed features according to knowledge and experiences of experts to describe the lung tissues. The advantages of the hand-crafted features were able to solve problems quickly. However, the disadvantages were obvious:since we had to design features manually, the designed features may be not suitable to handle other problems in different domains; the designed features were also incomplete, several features were needed to describe the objects; since the numerous feature parameters, the optimization progress was time-consuming; because feature extraction stage and classification stage were separated, parameters in both stages cannot be jointly optimized toward higher classification performance. Feature learning is a type of methods that draw useful and distinctive features from raw data automatically. Except the structural model parameters, there is no need to specify other parameters of these methods, therefore it is easy to apply them to solve different problems in different domains; as the prevalence of big data, traditional hand-crafted features cannot take advantage of the benefits of big data, whereas, based on the large model capacity, feature learning can accomplish this goal; some feature learning methods, for example, deep learning, is able to jointly optimize the feature extraction stage and classification stage.CNN (Convolutional Neural Network) is a type of feature learning methods, which has achieved great performance in computer vision. Considering the weaknesses of hand-crafted features and the unsupervised way of RBM (Restricted Boltzmann Machine), CNN is applied as the feature learning model. CNN can automatically learn discriminative features from lung tissue patches in a supervised way. In order to further promote classification performance, RF (Random Forest) is adopted to classify different types of lung tissues based on the extracted CNN features. RF is an ensemble method. It is not easy to overfit and is able to achieve better performance. A common phenomenon is also noticed among the previous works on the lung tissue classification problem. The imbalanced data distribution of different lung tissues and if the classifier is sensitive to the imbalanced data, the classifier would perform badly on the minority class. Instead of other data-level methods that aim to solve this problem, such as oversampling and SMOTE (Synthetic Minority Over-sampling Technique), this problem is solved in image patch preparation step. By changing the overlapping size of adjacent image patches, a relatively balanced data distribution is prepared for lung tissue classification problem.Despite that the CNN-RF model has achieved relatively good performance, there are still some drawbacks:the limited single-scale features and rotation-variant problems in lung tissues. We expect to learn multi-scale features which would provide richer information to describe the lung tissues and rotation-invariant features to accomodate the rotational variances in various lung tissue patterns. Motivated by the great success of Gabor-LBP (Local Binary Pattern) features to characterize the lung tissues, multi-scale and rotation-invariant Gabor-LBP images are used as the inputs of CNN, namely MRCNN (Multi-scale and Rotation-invariant Convolutional Neural Network). Based on the multi-scale and rotation-invariant features learned by CNN, the significantly higher classification performance is expected on the lung tissue classification problem.