Suspected Abnormal Cervical Nucleus Screening Based on a Two-Cascade Classifier

The accuracy of traditional computer-assisted cervical cancer screening systems often rely on fine cell image segmentation. In order to avoid the influence of segmentation accuracy on classification accuracy, a complete framework from a whole Pap smear image to screening results for suspected abnormal nucleus screening is proposed, which omits the traditional cell fine segmentation step. The framework is based on a two-cascade classifier, of which the first classifier aims to achieve ROI screening and the second one is used for nucleus classification. Firstly, the regions of interest (ROIs) in the cell clusters are extracted by adaptive thresholding and Selective Search to establish databases. Then, a new feature named Cshape is defined, which is combined with HOG features to train the first classifier used to screen out nucleus ROIs. Finally, the second classifier is trained by the Gray Level Co-occurrence Matrix, which is utilized to divide nucleus into suspected abnormal nucleus and normal nucleus, thus completing cervical cancer screening. Test results show that the highest accuracy of the two-cascade classifier is up to 100% and the random experiments verify the practicability of the framework.