ACCURACY OF NUCLEAR CLASSIFICATION IN CERVICAL SMEAR IMAGES - QUANTITATIVE IMPACT OF COMPUTATIONAL DECONVOLUTION AND 3-D FEATURE COMPUTATION

OBJECTIVE: To investigate the accuracy with which the nuclei of cells in overlapped and thick clusters in cervical/vaginal smears can be cla ssified independent of the segmentation algorithm used and to determin e the influence of three-dimensional (3-D) processing as compared to t o two-dimensional (2-D) methods on classification of the nuclei. STUDY DESIGN: Cell clusters were imaged from 31 ThinPrep smears composed of 808 nuclei, of which 420 were determined to be normal by a cytotechno logist. Sets of 2-D and 3-D volumetric features of the detected nuclei were formulated, and classifiers were constructed. The effect of comp utational deconvolution on classification was assessed using nearest-n eighbor and Wiener filter in 2-D and 3-D before calculating features. A ''best focus plane'' was calculated for each nucleus from the 3-D da ta set, and the 2-D features in this plane were also analyzed. RESULTS : The linear discriminant classifier operating on the 3-D features cor rectly classified 86% of the nuclei in clusters. Classification using the 2-D features in the best focus plane was correct for 85% of nuclei . Wiener filter deconvolution improved the classification accuracy by 3.2% and also changed the relative importance of the features contribu ting to the classification. The most influential features in the absen ce of deconvolution were geometric features, but when deconvolution wa s used, these changed to intensity-based features. CONCLUSION: Automat ed processing of cervical smears can be improved by accessing thick an d overlapped regions that are currently not processed by 2-D systems. For this, the thick regions should be selectively imaged in 3-D using standard transmitted-light bright-field microscopy, segmented as descr ibed in the companion paper, followed by classification using 2-D feat ures from the best focal plane after deconvolving only that slice from the 3-D image. Finally more sophisticated deconvolution algorithms ma y improve performance further.