A Study On Focal Liver Lesion Classification Based On Multiphase 3D CT Images

With the development of medical device and imaging technology,digitized medical image data also show a massive growth trend.In recent years,the rapid development of artificial intelligence is expected to help doctors reduce their work load,improve work efficiency,and gradually make people move toward smart healthcare.Cancer is the leading cause of death in the world.Among them,liver cancer ranks fourth.Early diagnosis and treatment are the most effective ways to reduce the mortality of liver cancer.However,due to various morphological changes of focal liver lesions,different types may have same features,which makes the task of classification of CT images of focal liver lesions a challenge and also a hot research topic in recent years.Due to the difficulty of acquiring labeled CT images of focal liver lesion and the lack of high-quality annotation data,the early related literature mainly focuses on the study of single phase and 2D typical layers.The well-designed manual features mainly focus on gray value and texture aspects.In recent years,various improved models of BoVW(Bag of Visual Words)are widely used in multiphase CT images and show good performance.However,the existing models can not be directly used in clinicalapplication because of the high requirement on the accuracy of medical image classification.Therefore,in this paper,the classification of focal liver lesions based on multiphase 3D CT images is studied.We focus on the key points such as feature extraction and selection and also the selection of classification models in order to improve the classification accuracy and efficiency.The main contribution are listed as follows.1)Propose a focal liver lesion classification algorithm based on the fusion of hand-crafted featuresThe traditional feature extraction algorithm can not well adapt to the three-dimensional and multiphase features of focal liver lesion CT images.The hand-crafted features often needs to be based on radiologists’ clinical diagnosis experience.In response to these problems,we focus on the evolution pattern of multiphase 3D CT images and propose a classification algorithm of focal liver lesion based on the fusion of various features.The algorithm extracts the hand-crafted features according to the bottom-up greedy strategy in terms of gray value,texture and shape,and selects the valid features based on the chi-square selection of the combined features.The experimental results show that this feature fusion and selection method significantly improves the classification performance,the final classification accuracy can reach about 75%.2)Propose a Bi-gram BoSTW model that combines N-gram and BoVW modelsBoVW model is a popular technique in the field of image classification.It can integrate shallow manual features and possess learning abilities.To some extent,the"semantic gap" can be filled,and the extensibility is strong.Based on this,we further propose a Bi-gram BoSTW model that combines N-gram and BoVW models.The temporal information of multiphase images is considered using a three-channel RGB representation.Meanwhile,the model introduces the N-gram technique in text categorization,which combines the spatial and texture co-occurrence information of visual words.The experimental results show that the method which not only combines the temporal information of multiphase images but also considers global distribution and local co-occurrence information of the lesion is very effective.The accuracy performance can reach 83%and its training speed is fast,which makes the model obviously better than the existing BoVW models and its variants.At the same time,the model can be generalized to other fields of radiology.Finally,we designed and implemented a CAD prototype system based on our method,which is suitable for the display,retrieval and classification of focal liver lesion CT images.