| The current gold standard of cervical cancer diagnosis is the histopathological diagnosis using tissues that are obtained by a doctor under the image guidance provided by a conventional white-light colposcopy.Hence the sampling accuracy of tissue directly affects the diagnosis accuracy.However,the imaging mode of a conventional colposcopy is relatively simple,and the lesion area is not highlighted,its value in guiding the doctor to determine the abnormal area is yet need to be improved.Therefore,this study proposes a computer-assisted diagnosis method that is to be jointly used with a narrow-band multispectral colposcopy.Main content of the research includes the optimization of the imaging device,the design and development of the humancomputer interface of the spectral colposcopy system,and multispectral pathological diagnosis algorithm.The spectral colposcopy system used in this study,uses circular arranged multi band LEDs as its illuminating light source,uses a high-resolution color CMOS camera as its imaging device.It can quickly capture,in a sequential manner,three images of the cervical tissue under the illumination of different wave band LED light sources,namely,regular white-light,narrow blue band,and narrow-green band,respectively.The multispectral pathological diagnosis algorithm developed in this study consists of two parts: image pre-processing procedures and tissue classification procedures.The image preprocessing procedures include the following steps: first step is filtering.In order to eliminate the noise on the original image,the median filter is first adopted to eliminate some possible extreme noises.The second step is registration to correct potential position offset among the tree images obtained at slightly different time,mutual information is used as the similarity measure function,and the gradient descent method is used as the optimization algorithm.The third step is the cervical tissue extraction.In this procedure,canny edge detection combined with morphological operation was used to extract the cervical tissue region of interest.The last step of preprocessing is to fuse multiple images,the objective of fusing is to maximize the contrast between the diseased tissues and normal tissue.This step enlarges the difference,or the "distance" between the abnormal and normal,hence facilitates the tissue classification procedures.The tissue classification procedures are applied to the multiband fused image.Its core concept is the K-means clustering algorithm of dichotomy(normal vs abnormal),and the initial seed value are the means of the sample points manually extracted by doctors in the sample graph.A contour coefficient method is applied in order to eliminate the specular reflection points of the cervical tissue.The color space component of HSI and Lab are also employed as the classification feature,in order to remove the shadow area which may otherwise affect the accuracy of the K-means clustering classification method.In this study,a preliminary clinical evaluation experiment is also conducted to validate the developed multispectral optical biopsy method.The digital classification results were compared with the results of the hospital pathological diagnosis.Experimental results indicate that the developed method achieves a sensitivity of 84%,a specificity of 93%,and an accuracy of 89%,which is rather high compared to what colposcopy experts can normally achieve with a conventional white-light colposcopy.The major contribution of this study is the development of a multispectral pathological diagnosis algorithm that is designed primarily for a reflective narrow-band colposcopy.This study has the potential to enhance the guiding value of a colposcopy to accurate tissue sampling.It can help to improve the accuracy of cervical cancer screening performed in hospitals with low-resources.It has practical value to improve the coverage rate of the early cervical cancer screening for women at the grass-roots level. |
PDF Full Text Request