Research And Application Of Lesion Detection Methods Based On Double Modules And Multi-modal Mechanism For CT Image

Deep learning can identify effectively lesion regions and improve the detection accuracy among a number of objective detection algorithms.Nowadays,the objective detection methods based on the deep learning only choose a certain tissue or organ for detection according the medical images,but ignore the correlations of various related sites on the characteristics of density,brightness and texture.Meanwhile,different stages of lesion areas also have a great difference.Thus,the thesis applies the double mechanisms for better feature extraction of lesion areas and the multi-module method to combine the image information and text information,and designs the lesion detection methods based on double mechanisms and multi-modal mechanism for CT images and related lesion detection system.The specific work is listed in the following.1)A feature extraction method of lesion detection based on double mechanisms for CT image is proposed to improve the ability of the discrimination between background and lesion regions during the process of lesion detection.Firstly,a double mechanism based on the attention module and global context module is introduced into the traditional YOLOv3 feature extraction network to extract fully lesion features and improve the ability of the network to distinguish lesion regions and background regions.Secondly,the high-resolution feature maps of the lower level is down-sampled and is added with the feature maps of the relatively higher level to make full use of the small lesion information in the high-resolution feature maps during the multi-scale fusion process.Finally,experimental comparisons are conducted with several related detection methods on the Deep Lesion public dataset and found that the method can achieve a good performance on the lesion detection task.2)A multi-modal liver lesion detection method based on a feature interleaving module is proposed to exploit fully effective features and apply clinical data to the process of lesion detection.First of all,the feature interleaving module is added to the YOLOv3 backbone network to extract lesion features.Then,the BERT network is introduced to extract features from the text and fuse the extracted text feature information with the image feature information.Finally,the GIo U position loss function is incorporated into setting a specific anchor frame for the dataset of this paper to finally detect the location of the liver lesion.The experiments show that the proposed method can fully exploit the lesion features based on the samples.3)A multi-site lesion detection system is further designed and implemented to assist the physician to better lesion detection.The system adopts the proposed lesion detection methods and includes login module,detection module and management module.The system satisfies the requirements of physicians and achieves a good detection accuracy in the process of multi-site lesions.