Deep Learning for Cervical Cancer Screening.

Abstract

Cervical cancer is the malignant tumor of the cervix. Cervical cancer remains one of the gravest threats to women world wide. As cervical cancer is a highly preventable disease, therefore, early screening represents the most effective strategy to reduce on the global burden of cervical cancer. Commonly used cervical cancer screening methods are pap smear, visual inspection with acetic acid, and colposcopy; however, these methods rely on the clincian’s visual inspection for diagnosis. Due to scarce awareness, lack of access to medical centers, and highly expensive procedure procedures in developing countries, the vulnerable patient population cannot afford to undergo examination regularly. Additionally, there is small number of oncologists in developing countries in relation to cancer cases. The application of machine learning more specifically computer vision in the medical field has proven to be less tedious, less time consuming and provides accurate diagnosis. Some of the use cases of computer vision in the medical field are radiology, oncology, dermatology and laboratory test automation. Several computer vision based studies have been done to aid the screening of cervical cancer; Ashutosh Kanitkar [3] proposed a novel deep learning approach to perform binary and multiclass classification for cervical abnormality detection however; this was limited by small data and no deployment was done. In this research project, we implemented a deep learning model to aid in the screening of cervical cancer. In addition to the cervical cancer screening task, we also implemented a deep learning model to detect and classify various cervix types that is; Type 1, Type 2 and Type 3. We achieved an mAP of 0.879 for the cervical cancer screening task and an mAP of 0.646 for the cervix type detection task. We were able to deploy the models for both tasks unto an Android smartphone.