Classification Of Benign-Malignant Nodules Based On Deep Learning And Multi-Feature Fusion

The latest research results released by the International Agency for Research on Cancer（IARC）indicate that lung cancer is the second most commonly diagnosed cancer globally.The physicians mainly determine the benignity and malignancy of nodules in clinical diagnosis by analyzing the CT images of patients,while the malignant-nodule is the main cause of lung cancer.Earlier detection and identification of malignant lung nodules can lead to better treatment and improved survival of the patients with lung cancer.However,the rate of patient misdiagnosis and missed diagnoses cannot be underestimated due to the influence of manual film reading and subjective physician experience.The emergence of computer-aided diagnosis systems based on machine learning have largely improved the effectiveness and accuracy of doctors’diagnosis and provides for efficient diagnosis and treatment of patients.It has been shown that visual features of lung nodules,such as subtlety,spiculate and calcification play an important role in the diagnosis of malignancy.However,the gap between attributes and deep features is the main factor limiting the performance of computer-aided diagnosis.Therefore,this thesis investigates a method for classifying benign-malignant lung nodules based on deep learning and multi-feature fusion.The major work is as follows:The integrated learning classification model based on attributes and deep learning is proposed to deal with the problem of utilizing attribute features in the processing of nodule benign and malignant classification.Firstly,the attribute features are obtained from clinical information while the deep features of nodules are extracted from the preprocessed computed tomography（CT）images.Secondly,the Fuse-Convolutional Neural Network（F-CNN）model is proposed for highlighting the essential role of attributes in the classification processing which integrates deep features and attribute features mapped through the transposed convolution.Meanwhile,the Fuse-Long Short-Term Memory（F-LSTM）model is proposed to focus on specific deep features for classification via the affine transformation of attribute features.Finally,the prediction scores of the F-LSTM and F-CNN models are fused by an integrated learning algorithm for early identification of malignant lung nodules.The model is validated on a publicly available lung nodule dataset（LIDC-IDRI）,which not only demonstrates that attribute features have an important role in lung nodule classification.Moreover,the integrated learning algorithm can effectively improve the classification effect of the model.It provides a new direction for the development of an assisted diagnosis system for the classification of benign and malignant pulmonary nodules.The classification model of benign-malignant nodules based on uncertain labels and multi-feature fusion is proposed to address the problem of uncertain labels in the dataset.Firstly,the LIDC-IDRI dataset is divided into true reliable dataset（TR Set）and low reliable dataset（LR Set）according to the expert labeled dataset,and the dataset is expanded using a data augmentation algorithm.Secondly,the DF Block is proposed to perform nonlinear mapping and linear fusion of the deep features extracted from the TR Set and LR Set.Finally,the augmented features are fed into the F_DF-LSTM and F_DF-CNN models for feature classification,and the prediction results of the two models are fused as the ultimate prediction results using the weighted average method.The experimental results show that the model learns more helpful information by dividing the dataset and integrating the two parts of the features in a linear way.The classification performance is superior than the currently available methods for classifying benign and malignant lung nodules.In short,the lung nodule classification model based on deep learning and integration of multiple features proposed in this thesis focuses on the essential role of attribute features in the classification process.It also provides a new solution idea for the application of uncertainty labels,which enables the model to obtain a more excellent classification performance.