The Research Of Computer-aided Diagnosis In Chest Images Based On Multi-semantic Task And Multi-label Incremental Learning

Recently,the deep cross-fusion applications between artificial intelligence(AI)and medical imaging have attracted wide attention of many researchers with the con-tinuous growth of massive image data,continuously iterative update of intelligent di-agnostic model algorithms,substantial improvement of computing power and strong support of national policies.In the academia,artificial intelligence,especially deep learning,has made considerable progress on many long-term problems such as lesion detection,segmentation and diagnosis.The research on medical image-assisted diag-nosis shows a booming development.In the industry,AI medical imaging is rapidly transitioning from the experimental stage to the clinical trial stage with the develop-ment advantages of medical image big data and deep learning technology.AI medical imaging has become one the most closely integrated fields of artificial intelligence and medical industry.This dissertation focuses on the problems of imbalanced data distribution,high misdiagnosis rate and a small number of new diseases diagnosis in the chest imaging-assisted diagnosis scenarios.We carry out researches on generative adversarial learning-based data synthesis over-sampling,multi-order transfer learning and multi-label incre-mental learning.The main research work and contribution of this dissertation include the following four parts.1.WGAN-based synthetic over-sampling method for semantic fine-grained classification of pulmonary nodules in CT images.For the seriously imbalanced data distribution issue in the semantic fine-grained classification of pulmonary nodules,we attempt to synthesize samples by approximat-ing the distribution of raw data in minority classes by wasserstein generative adversar-ial network(WGAN).The WGAN-based synthetic over-sampling method is proposed for the sample synthesis supplementation of the semantic attributes'subclasses data of pulmonary nodules.The convolutional neural networks(CNN)are used for the fine-grained classification of seven semantic attributes of pulmonary nodules in CT images.The WGAN-based synthetic over-sampling technique can synthesize useful samples for the minority classes,which is beneficial to the training of CNN fine-grained classifica-tion models.WGAN-based scheme can achieve better fine-grained classification per-formance than the schemes of traditional data augmentation,typical GAN and DCGAN.It indicates that the WGAN-based scheme can effectively solve the serious imbalance problem in semantic fine-grained classification of pulmonary nodules.2.Auxiliary diagnosis of benign and malignant pulmonary nodules based on multi-order association transfer learningTo solve the problem of high misdiagnosis rate of benign and malignant pulmonary nodules,we propose a multi-order association transfer learning framework for multi-attribute pulmonary nodules to improve the accuracy of pathological benign and malig-nant diagnosis of pulmonary nodules.The experiments show that only three semantic tasks,i.e.,texture,diameter and lobulation,are needed as the source task of correlative transfer learning.It takes 26.99%of the labeled data to achieve the best performance in the benign and malignant diagnosis of pulmonary nodules.The method of multi-order association transfer learning proposed in this dissertation performs better than the existing work in the pathologically confirmed benign and malignant classification of pulmonary nodules.The classification results are equivalent to the judgment results of radiologists,and even exceed the subjective judgment of doctors in the accuracy index.3.Multi-order semantic task transfer graphs construction and quantitative analysisTo maximize performance gain for a set of semantic target tasks,we combine dif-ferent semantic tasks for multi-order association transfer learning to quantify the re-lationships between different semantic tasks.Multi-order semantic task transfer rela-tionship graphs under the constrain of different supervision budgets are constructed to match the best transfer source for different target tasks.Then we quantitatively ana-lyze the relationships between the annotation data quantity in supervised learning and the performance of the target task results and demonstrate the relationships between semantic annotations and the overall performance of all semantic target tasks.The Ex-perimental results show that it can achieve 91.3%of the semantic overall performance target tasks with 39.28%of semantic annotations and reach 96.6%of the semantic over-all performance target tasks with 56.63%of the amount of semantic annotation through the third-order transfer learning of semantic association.The quantitative trade-off be-tween semantic annotations and overall performance of the target task can provides a reference for the most cost-effective labeling.4.Auxiliary diagnosis of common thorax lesions based on multi-label incre-mental learningIn view of the diagnosis learning for a small number of new lesions,we propose a small-sample multi-label incremental learning framework.First,we carry out multi-label learning diagnosis for 14 common chest X-ray lesions such as pulmonary nodules,atelectasis,etc.Second,6 new lesions such as tuberculosis,etc.are incrementally diag-nosed.In order to improve the identification and diagnostic performance of multi-lesion new labels in chest X-ray,we derive and propose a MLSGM-based feature regulariza-tion prior.The experiments show that the MLSGM feature regularization is conducive to enhancing the generalization ability of depth feature characterization in new data samples.The proposed small-sample multi-label incremental learning program is fea-sible and practical in clinical practice for the diagnosis of thorax lesion abnormalities from the perspective of time and resources.Driven by the small-sample multi-label in-cremental learning framework,it is potential to realize the auxiliary diagnosis for more and more thorax lesions in clinical practice.To solve the medical problems in the small-sample field is more important for the development of artificial intelligence in medical imaging,because the cost of the col-lection and labeling of these medical data is very high and small-sample scenarios often occupy most of the actual scenes.The methods of the generative adversarial synthesis learning,multi-semantic task association transfer and multi-label incremental learning are the core to solve the problems of small data.The improvement of the accuracy and efficiency in medical image-assisted diagnosis plays an important role and practi-cal significance in alleviating medical resources,promoting graded diagnostic treatment services and promoting telemedicine services.