Automated Cervical Cell Segmentation using Residual Learning and Attention Mechanism for the Cervical Cancer Diagnosis

Cervical cancer, one of the common deadly gynaecological cancers among women, can be prevented by regular screening to detect the existence of any premalignant cervical cells at the beginning stages. One of the widely used screening tests for the early diagnosis of cervical malignancy is Pap smear test. Here cytopathologist identifies the malignant cells through the microscopic observation and it can lead to manual errors. Deep learning based automated cervical cancer diagnosis has the ability to reduce these errors and provide more accurate results. But the accuracy of this automated diagnosis depends a lot on the accurate segmented cytological images. Deep learning based cervical cytology image segmentation is still in its developing stage due to the unavailability of high quality cervical cytology image dataset without false negative object issue. In this work we are proposing the automatic cervical cell segmentation using two deep neural networks ResUNet and ResUNet with Attention on a new high quality cervical cytology dataset Cx22 which is free from false negative object issue. The proposed models provide better dice score values for nucleus semantic segmentation compared to the baseline models.