Automatic Classification of Whole Slide Pap Smear Images Using CNN With PCA Based Feature Interpretation

Abstract

Classification of whole slide image (WSI) cervical cell clusters traditionally involved two stages including segmentation to crop single cell patches followed by the classification of single cell patches. Hence the performance of classification pipeline depends on segmentation accuracy. We propose a first-time-right method which is a segmentation-free direct classification of WSI cervical cell clusters (without the extraction of single cell patches). The proposed method is evaluated on SIPaKMeD and Herlev datasets. Our method significantly outperformed previous methods and baselines with an accuracy of 96.37% on WSI patches (cell clusters) and 99.63% on single cell images. We also propose a PCA based feature interpretation method to visualize and understand the model to make its decisions more transparent. Our solution is promising in the development of automatic whole slide pap smear image classification system.