Automated image analysis in multispectral system for cervical cancer diagnostic

Abstract

Uterine cervical cancer is the second most common cancer in women worldwide. The accuracy of colposcopy is highly dependent on the physicians individual skills. In expert hands, colposcopy has been reported to have a high sensitivity (96%) and a low specificity (48%) when differentiating abnormal tissues. This leads to a significant interest to activities aimed at the new diagnostic systems and new automatic methods of coloposcopic images analysis development. The presented paper is devoted to developing method based on analyses fluorescents images obtained with different excitation wavelength. The sets of images were obtained in clinic by multispectral colposcope LuxCol. The images for one patient includes: images obtained with white light illumination and with polarized white light; fluorescence image obtained by excitation at wavelength of 360nm, 390nm, 430nm and 390nm with 635 nm laser. Our approach involves images acquisition, image processing, features extraction, selection of the most informative features and the most informative image types, classification and pathology map creation. The result of proposed method is the pathology map - the image of cervix shattered on the areas with the definite diagnosis such as norm, CNI (chronic nonspecific inflammation), CIN(cervical intraepithelial neoplasia). The obtained result on the border CNI/CIN sensitivity is 0.85, the specificity is 0.78. Proposed algorithms gives possibility to obtain correct differential pathology map with probability 0.8. Obtained results and classification task characteristics shown possibility of practical application pathology map based on fluorescents images.

Keywords

• medical image processing
• computer vision
• Clinical decision support systems
• machine learning