Algorithm Research On Bone And Tissue Segmentation From Chest Radiograph And Automatic Measurement

Bone segmentation and lung field segmentation on X-ray chest radiographs are of great value for the automatic diagnosis and detection of lung diseases on X-ray chest radiographs.But the low contrast of X-ray chest radiographs and the image overlap between tissue structures result in traditional methods that cannot achieve good results.According to the different features and difficulties of bone segmentation and lung field segmentation on X-ray chest radiographs,we propose two different semantic segmentation networks to accomplish the two tasks and a variety of automatic measurement methods of medical indicators.The main tasks as follows:(1)A clavicle and rib segmentation dataset containing 88 X-ray chest radiographs was constructed.Each X-ray chest radiograph has four mask pictures of its corresponding clavicle,anterior rib,posterior rib,and all bones(clavicle + rib).(2)A densely connected multi-task semantic segmentation network is proposed for segmentation of the clavicle and ribs on X-ray chest radiographs.Firstly,aiming at the problem that the features cannot be fully explored due to the small dataset,the method of feature reuse and feature transfer is adopted.Secondly,the multi-label problem is transformed into a multi-task problem and a feature separation network is designed to complete different tasks at the same time to reduce the time cost.Finally,a mask coding mechanism is proposed,which at the same time avoids the poor segmentation effect caused by selecting improper threshold.Experiments were performed on our dataset.The DSC values of this method on the four tasks of clavicle segmentation,anterior rib segmentation,posterior rib segmentation,and all bone segmentation reached 93.78%,80.95%,89.06%,and 88.38%,respectively.(3)A semantic segmentation network based on multi-scale convolution and feature pyramid is proposed to segment lung field on X-ray chest radiographs.In the network,multi-scale convolution and multi-scale feature fusion based on feature pyramid are used to improve the situation of under-segmentation and over-segmentation in network prediction results.And a loss function that considers both the accuracy of a single pixel and the performanceof the entire image segmentation is proposed to help network obtain the optimal parameter configuration.This method is validated on the Montgomery dataset,and the best result is obtained.(4)A variety of automated measurement methods for medical indicators are proposed.Based on bone segmentation and lung field segmentation on chest radiographs,automatic measurement methods of medical indicators such as the costal angle and cardiothoracic ratio is proposed.We propose two neural networks to complete bone segmentation and lung field segmentation on X-ray chest radiographs,respectively,and propose automatic measurement methods for various medical indicators.This not only reduces the workload of the doctor,but also is an important part of the computer-aided diagnosis system.