Abstract:
Cervical cancer is the most common cancer among women worldwide. It is screened using visual
inspection after application of acetic acid (VIA), Papanicolaou (Pap), and Human papillomavirus
(HPV) test and diagnosed by histopathology test. Currently, identification of cervix type and
histopathologic image classification is examined manually, thus big inter-and intra- observer
variability exists. Moreover, incidence and mortality rates are consistently increasing in
developing countries due to a shortage of screening facilities, skilled professionals, and lack of
awareness. Therefore, there is a need for an automated system that overcomes subjectivity and
inconsistency in the screening process. Different studies have proposed the classification of cervix
type and cervical cancer using deep learning and machine learning techniques. However, most of
the proposed techniques in the literature are limited to specific class classification. This study aims
to develop an automatic cervix type and cervical cancer classification using deep learning
techniques. For cervix type classification 3872 colposcopy images were collected from the Kaggle
dataset. In addition, 133 colposcopy images and 915 histopathologic images were collected from
local health facilities for cervix and cervical cancer classification from local health facilities.
Different pre-trained models including VGGNet, ResNet, Ensemble of MobileNet, and ResNet
and EfficientNet were trained and compared for cervix type classification. Prior to classification,
a lightweight MobileNetv2-YOLOv3 model was created for the region of interest (ROI) extraction.
The extracted cervix images were then fed to the effecientnet_b0 model for cervix type
classification. A mean average precision (MAP) of 99.88% for the region of interest (ROI)
extraction, an accuracy of 97%, and a Kappa score of 0.95 for the cervix type classification task
were achieved using the selected models. For cervical cancer classification, an efficient_b0 pre trained model was trained and validated using histogram matched histopathological images, and
an accuracy of 95% and Kappa score of 0.92 were achieved. Even though the best performing
model was constructed further improvement needs to be done using more data and by classifying
sub-class of histopathology images to know disease prognosis. In addition, to make full cervical
cancer screening and diagnosis method cytology image classification model construction is
required
