Bone Edge Detection And Bone Suppression Of Chest Radiographs Based On Convolutional Neural Networks

Chest radiography is one of the most common and cost-effective medical imaging examines for detecting lung findings and lung pathologies in the clinical setting.However,as chest radiographs are two-dimensional projections of the three-dimensional human body,overlapping anatomical structures such as the ribs and clavicles obscure the soft-tissue components within lung areas,making it difficult for radiologists to read and computer-aided detection and diagnosis systems to analyze CXRs.Accurate location of bone structures and suppression of bone components in CXRs are of clinical importance for radiologists and CAD systems to analyze diagnostic chest images.However,existing image processing methods often neglect to introduce anatomical a priori into chest films and generate bone suppression images that often differ in style from clinically meaningful DES images.The research objective of this paper is bone edge detection and bone suppression of chest X-rays(CXRs)based on convolutional neural network.The main research work of this paper is as follows.(1)Automatic delineation of ribs and clavicles in CXRs using fully convolutional DenseNets(FC-DenseNets).Unlike previous methods in which multiple complex image processing steps are used to detect bone edges of CXRs,we train a deep model for automatic and fast delineation of ribs and clavicles in CXRs using an end-to-end FC-DenseNet.Therein,we propose a pixel-level weighted loss function for the uncertainty of manual outlining to reduce the interference of uncertain pixels to the network training.The proposed method can handle various CXRs from different devices and automatically locate their bone edges such as ribs and clavicles.(2)Bone-Edge-Guided Generative Adversarial Network for Generating Dual-Energy Subtraction Soft-Tissue Images from Chest Radiographs(EGAN).Based on the work of detecting the bone edges of CXRs,the EGAN model takes the bone edge information of CXRs as a priori input to the network and guide the network to suppress the bone components more efficiently and accurately.Specifically,the CXR and its bone edge probability map were used as input to train the bone suppression model using GAN as the basic network.Experiments on our collection of 504 DES data show that the proposed EGAN can produce soft tissue images that are highly similar to real DES soft tissue images.(3)Application on Tuberculosis Recognition of Task-specific Feature guided Bone Suppression Model.In this work,different from the edge-guided bone suppression model that aims to generate high-quality soft tissue images,we propose a task-specific feature-guided bone suppression model aiming to generate more discriminative soft tissue images for improving TB recognition performance.We evaluated the impact of bone suppression on the tuberculosis recognition by training the respective classification models for triple classification(healthy/sick but non-TB/TB)of TB CXRs using the chest X-ray images before and after bone suppression.Experimental results indicate that the TB recognition model trained on the bone-suppressed images generated from our BS model achieves superior performance to the state-of-the-art methods.In this paper,we systematically studied the bone suppression method of CXRs from three aspects:bone edge detection of CXRs,bone edge guidance,and task-specific feature guidance.We proposed automatic bone edge detection model based on FC-DenseNets,edge-guided bone suppression model,and task-specific feature-guided bone suppression model and verified the effectiveness of bone suppression images for computer-aided tuberculosis diagnosis.