Detection and Classification of Cervical Exfoliated Cells Based on Faster R-CNN
Du, Xueyu Li, Qinghua Li · 2019
The detection and classification of cervical cells via Pap smear or Liquid-based cytology (LBC) have important clinical significance for pathological diagnosis. For the limitations of pathological images and the complexity of morphological changes of cells, the accurate classification of cervical cells remains a challenging task. DNA ploidy analysis can complete the classification and recognition of individual cells well, and misclassification is common because of the factors of inevitable imaging. Unlike traditional classification methods that rely on manual features, we proposed a region detection and classification method based on multi-semantic label combined with morphological information analysis. We followed the principle that a nucleus represents a cell and divided cells into five categories. The overlapping cell regions could be divided into multiple single cells. In this paper, the data pre-processing was made up for the imbalance in classification. The weight parameters obtained from ImageNet dataset pre-training were realized the transfer learning of cervical exfoliated cells based on Faster R-CNN on LBC dataset. The result showed that the mean average precision (mAP) was 66.98% and the accuracy was 91.61% by learning depth and superficial features of cervical exfoliated cells with ResNet-101 as the backbone. To some extent, it can assist pathologists in the early diagnosis of cervical cancer.