Research And Application Of CT-image Lesion Detection Based On YOLOv3

Object detection is an indispensable essential and primary task in the field of Computer Vision,and it is of high research value and significance.In recent years,the morbidity and mortality of cancer have shown a trend of sharp increase year by year.The lesion areas are mainly concentrated in the various organs and tissues of the chest and abdomen,which,however,are the primary guarantee for maintaining human life and health.Therefore,it is urgent to study a high-precision lesion areas detection method and put it into practice,which has far-reaching significance for assisting early imaging screening,reducing the rate of radiologists’ misdiagnosis and the risk of cancer.However,the particularity of medical imaging itself brings huge challenges to lesion areas detection.Due to the relatively high spatial resolution,high screening efficiency and costeffectiveness,CT images are widely adopted and recognized by physicians and radiologists.Taking the chest and abdomen CT images as the research object,and YOLOv3 method as the baseline,this thesis conducts research on the detection of lesion areas locating on liver and chest-abdomen multi-site,respectively.The main research contents of this thesis are as follows:(1)Aiming at the problem of small-scaled dataset and relatively small-sized lesion areas,a CT liver lesion detection method based on training sample expansion and multilevel prediction is proposed.This method employs a specific color-space data augmentation method,in the field of medical imaging,based on fixed CT window levels and varying window widths to expand the training samples.For the problem of relatively small-sized lesion areas,the multi-level predictions of detection network is adjusted,and more anchors are set on the feature maps with larger resolution.Meanwhile,to deal with the problem of the changeable position of lesion areas,an excellent method is selected as baseline;Experiments show that this method can effectively alleviate the phenomenon of missed detection and false detection.(2)Aiming at the problem of insufficient feature extraction capability of the detection backbone and the more complex and changeable location of lesion areas in new dataset,a method of multi-site lesion detection in chest and abdomen CT based on attention map and feature enhancement is proposed.This method employs the hourglass backbone network to enhance feature extraction,which improves the discriminating capability of targets’ location;The collocation of two embedded modules further enhances and fine-tunes the feature representation.Meanwhile,to deal with the particularity of the prediction form in single-class detection task,the multi-class classification scores prediction in general detection tasks is detached to adapt singleclass detection;Experiments conducted on the public Deep Lesion dataset prove the effectiveness of the improvement and the superiority of the proposed method.(3)The chest-abdomen CT-images lesion areas detection system is designed and developed.After fully understanding the actual needs of radiologists,three modules are subdivided to realize the functions of preprocessing,object detection and postprocessing optimizations,respectively.The system puts the finally proposed detection model into practical application,and effectively alleviates the heavy burden in radiologists’ daily work-flow,promoting the development of computer-aided diagnosis and even the entire smart medical care.