| Gastrointestinal cancers have a high incidence and mortality rate in China.They become a huge economic and medical burden on the nation and seriously affect the health of the people nationwide.Early detection and treatment is crucial to improve the five-year survival rate of patients with gastrointestinal cancers.So early screening of tumors or precancerous lesions is important to improve the health of people.Confocal laser endomicroscopy(CLE)is a fluorescent microendoscopy system based on the principle of confocal imaging,with ultra-high magnification and spatial resolution.Real-time analysis of the microstructure of the digestive tract mucosa can be carried out using this device during examinations,which helps to find mucosal lesions that are not obvious.This helps physicians to make an early and correct diagnosis of disease while minimizing the discomfort for patients.The high-speed scanning system and the imaging fiber bundle of the resonant mirrors can improve the imaging speed and make the end of CLE more flexible.This helps to improve image resolution and operation flexibility,making it more clinically effective.But how to maintain the imaging quality of CLE while improving imaging speed and flexibility is a challenge.There are many factors affecting the imaging quality,such as dislocation caused by progressive scanning,distortion caused by speed variation of resonant mirrors,and structured noise caused by heterogeneous light transmission inside the fiber bundle.This paper focuses on a high-frame-rate optical-fiber beam confocal endomicroscopic imaging system and detailedly discusses the methods to solve the above image quality problems.This study develops a confocal endomicroscopic imaging system that uses fiber bundles to deliver endoscopy images and resonant scanner to scan progressively,to carry out in vivo fluorescence imaging with high speed and resolution.The main components of the system,control system,data collection and imaging software are discussed in detail in this paper.The scanning system uses resonant scanner with a frequency of 8 KHz,and an imaging frame rate of 30 fps can be realized when images are acquired using progressive scanning.An FPGA-based image acquisition system was designed in this study,and the parallel computing methods and tools for FPGA were used to filter,align,and correct the image data.User Datagram Protocol(UDP)with low latency was used to transmit data.And integrity verification and packet loss fixing mechanisms were used during the image transmission to improve the reliability of communication.Practical video display and image management software was developed and can be used on the computer.The built system underwent clinical trials,which demonstrated its efficacy in accurately differentiating between normal and cancerous mucosa.This study also investigate the imaging quality problems such as aberrations,dislocation and fiber structured noise when the confocal microendoscopic imaging system was used.We analyzed the causes of image aberrations and dislocation and provided an FPGA-based image correction method.The correction formula of sampling time was derived based on the motion law of resonant scanner,and the image aberrations were corrected by changing the sampling interval.For the problem of dislocated odd and even lines of the image,a global correction was performed using a genetic algorithm.A line-by-line correction method was designed to correct the residual local dislocation.The proposed method has successfully rectified the aspect ratio of the fiber bundle from 0.96 to 1.0,and reduced adjacent row misalignment to 0.23 pixels.For the structured noise in fiber bundle images,the Delaunay triangulation technique was employed to partition the fiber bundle image into dense triangular tessellations,followed by reconstruction of the fiber bundle image via the center of gravity interpolation method.Then,an objective function of fiber bundle image restoration based on total variation regularization method was created.The structured noise in the fiber bundle images was removed through a two-step iterative shrinkage thresholding algorithm.This paper also investigates the method of removing structured noise from opticalfiber images by deep learning,designs a generative adversarial network and introduces an attention mechanism.The greater weight of the fiber core in the attention graph was leveraged to make the generative adversarial network focus more on the fiber core informtion.Training data were created to train the generative network and the discriminative network,and the network model was tested using synthetic fiber images and real fiber images.The results showed that the optical-fiber structure in the images could be effectively removed by the attention-based generative adversarial network.In conclusion,this paper presents the development of a high-frame-rate CLE system,and investigates image distortion,dislocation,and fiber bundle structure issues,resulting in the acquisition of high-quality images. |
PDF Full Text Request