Research On Lung Nodule Recognition Algorithm Based On Deep Learning

Computed tomography(CT)technology is an important tool for detecting early stage lung cancer.The Computer Aided Detection(CAD)system for lung is a comprehensive application of medical image processing and machine learning technology,which is designed to rapidly and accurately detect nodules and identify the benignity and malignancy of lung nodules from CT,thus providing an efficient assisted diagnosis and treatment plan.Many machine learning algorithms have been applied to lung CAD systems,mainly divided into traditional machine learning methods and deep learning,each with its own strengths.Traditional machine learning methods have a strong theoretical foundation,while deep learning algorithms have more efficient processing capabilities,so specific solutions need to be proposed for specific problems.To improve the feature extraction and recognition module of lung CAD system,we designed and constructed the Multiple Kernel Learning Support Vector Machine(MKL-SVM)algorithm to identify lung nodules based on deep fusion features and swarm intelligence optimization.Deep learning algorithms are combined with improved traditional algorithms to exploit the advantages of both and improve the accuracy of nodule identification and reduce the occurrence of false positives and false negatives,mainly including:1.The classical deep learning network VGG16 is used for deep feature extraction of the region of interest(ROI)of candidate nodules in lung CAD,and the deep features are combined with handcrafted features as the final feature vector to take into account more comprehensive feature information of nodules.On the one hand,it avoids the disadvantage that the potential information of nodules cannot be obtained by a single handcrafted feature,and on the other hand,it complements the lack of interpretability of deep features.2.Designing appropriate kernel functions for Support Vector Machine(SVM)for a specific problem is still a difficult task.Therefore,an improved MKL-SVM algorithm is proposed for the lung nodule identification problem,which constructs a multiple kernel function in a linear convex form of the polynomial kernel with high generalization ability and the sigmoid kernel with high learning ability,so as to avoid over-fitting during the model training process and increase the model generalization ability.3.To further prevent the missed detection of nodules and to take into account the overall recognition accuracy,the accuracy and sensitivity are introduced into the concept of harmonic mean in statistics,a new score function named F-new is used as the evaluation criterion for the subsequent recognition results,and its feasibility is analyzed and used as the fitness function for the subsequent parameter search algorithm.Compared with the single accuracy or sensitivity as the fitness function,the F-new function can improve both the detection rate of lung nodules and the overall recognition effect,and achieve multi-objective optimization.4.To overcome the problems of long training time and complex search process of grid search method,the swarm intelligence strategy is introduced for parameter search to shorten the training time of the model.At the same time,in order to overcome a lack of particles diversity in the early iterations of the Particle Swarm Optimization(PSO)algorithm,a hybrid swarm intelligence optimization strategy is suggested.By introducing the Simulated Annealing(SA)algorithm of global optimization to help the particles jump out of the local optimum and better seek the global optimum solution,the hybrid SAPSO optimization algorithm of SA and PSO is used as the parameter seeking method of MKL-SVM so as to improve the accuracy of the model to recognize lung nodules and reduce the missed detection.