3D CT Abdominal Organ Detection Based On HOG3D

Computed Tomography(CT)is a common and important auxiliary medical treatment in clinical medicine.It has important implications for doctors' diagnosis and decision making.With the development of medical imaging,intelligent automatic positioning,segmentation and other medical image processing methods have become more and more popular among people in industry.Among them,the detection and location problem is a prerequisite for many medical aids.The efficient and intelligent positioning method can improve the efficiency of the doctor's pathological analysis and make them more focused on treatment and diagnosis.Therefore,this paper presents an innovative automatic organ localization method in 3D CT images based on HOG3 D method.Firstly,in order to improve the quality of gradient information extracted from feature extraction and the noise in CT images,this paper proposes an improved GHOG3 D feature extraction method for HOG3 D.The advantage of the improved extraction method is that it can be achieved by using the Gaussian kernel convolution method.The effect of gradient calculation can also smooth the noise in the CT image,ensuring the stability of the gradient calculation.Secondly,this paper inspired by the classic 2D HOG feature pedestrian detection method,and proposes a set of 3D CT image detection method based on HOG3 D.For the training process,the sample data needs to be preprocessed and normalized firstly,which can ensure that the calculated HOG feature gradient is in the same order of magnitude.Then the various samples of the data set are unified to the same size and feature extraction is performed.By Gradient descent method training,the final detection template is trained for the organ to be detected.For the testing process,in order to ensure that the organs to be detected under different resolutions will not be missed,this paper constructs a feature pyramid and uses a trained model to detect each layer of the feature pyramid.Finally,it was verified on the sliver07 dataset and the CT images datasets acquired from 40 patients by expert annotation.The experiments showed that the HOG3 D features had a good effect on organ detection in 3D CT images.Thirdly,in order to improve the detection result more precisely,this paper proposes a multi-scale model detection method for severe changes in the size of abdominal organs.Firstly,the training samples are scaled up and down,and a number of organ detection models of different sizes are pre-trained for the scaled samples and detected on the CT images.The experiment finally showed that the detection results of the multi-scale model are more accurate than those of the single-scale model.At the same time,inspired by the RCNN region regression method,this paper proposes a 3D region regression method.The transform coefficient model is obtained by learning scaling and translation coefficients,and the learned transform coefficients can be applied to the initial detection model to achieve fine detection.Effect.Finally,re-testing on the data set showed that the detection results of various organs have improved to varying degrees.