Automatic Lung Tumor Segmentation Based On Convolutional Neural Network In PET CT Image

Accurate lung tumor segmentation plays an important role in the process of clinical diagnosis and treatment.However,due to the uneven distribution and uncertainty of lung tumors in space,the accurate segmentation of lung tumors has always been a serious challenge.At present,PET images and CT images are widely used in the diagnosis and treatment of lung tumors.In this paper,a novel method is presented for fully automatic segmentation of lung tumors on medical images.It can overcome above-mentioned problems and achieve full-automatically accurate segmentation of the patient's tumors in the patient's whole body medical images.The proposed algorithm is divided into two parts:segmentation of lung tumors on PET images based on convolutional networks composed of sparse feature maps and segmentation of lung tumors on PET-CT images based on feature fusion.In the first part,the trainable compressed sensing module CSM is used to join the fully convolutional network.It can realize information compression during the network training process,remove redundant feature maps and enhance effective feature maps.This module reduces the problem of high segmentation false positives caused by similar intensity of heart,spine,liver and lung tumors in PET images.Meanwhile,the deep supervision mechanism is used to constrain each trainable compressed sensing module and then train the weight parameters.It is aim to refine the effective features and transform the equal channel feature maps into discriminative feature maps.In the second part,the trainable independent component extraction module ICEM is used to obtain a set of independent feature maps from mixed PET-CT features.It can also obtain independent components related to lung tumors,and remove irrelevant independent components.By combining the proposed trainable independent component extraction module ICEM and the trainable compressed sensing module CSM,a dual-modal fusion module(Fusion)is formed.This dual-modal fusion module is embedded into the PET-CT dual-modal lung tumor segmentation network.The network can integrate the information of PET and CT images very well and such that the features are effectively extracted for lung tumor segmentation.The proposed algorithm is tested with 3D PET and CT images from 100 patients with non-small cell lung tumors.It is also compared to related mainstream deep learning image segmentation algorithms and lung tumor segmentation algorithms.The quantitative and qualitative experimental results show that the algorithm proposed in this paper has higher segmentation accuracy in terms of lung tumor segmentation performance.