Segmentation and Multi-Label Classification of Visual Cervical Pathology by Deep Neural Networks

The diagnostic capabilities of segmentation and multi-label classification methods by deep neural network (DNN) model are investigated for the visual evaluation of cervical pathology. A total of 8907 de-identified cervical images sourced from open data repositories were utilized. Employing a visual evaluation methodology, 2012 non-standardized cervical images were selectively extracted from the general dataset for further training within the DNN model developed on the basis of U-Net architecture with ResNet encoder and decoder components. The diagnostic performance of the DNN model was assessed using pertinent metrics, including accuracy, precision, recall, and F1-score for segmentation, as well as accuracy and area under receiver-operator curve (AUC) for multi-label classification. The quite small difference between precision, F1, and especially for recall values was observed for training and validation parts of the dataset for class 0 (absolute normal). It allows us to distinguish reliably the “absolute normal” cases from the others. The proposed multi-label classification approach can be used effectively for determining the cervical pathology, and the image segmentation approach within a single model allows us to determine the visual boundaries of cervical pathology, which can be of key importance when using this DNN model at the stage of pre-screening diagnostics. In general, further research will be targeted on increase of the relevant dataset with potentially much better performance of the diagnostic capabilities of the proposed DNN models that can be used as help tool in the current and future practices related to the assessment of cervical pathology in the context of cervical cancer problem.