Deep Learning for Cervical Cancer Lesion Segmentation in Mobile Colposcopy Images.
Abstract
Cervical cancer (CaCx) is a serious public health issue that mostly affects middle-aged women worldwide, particularly in less developed nations. Around 342 000 women died from cervical cancer worldwide in 2020, with an estimated 604 000 women receiving the diagnosis. Through HPV vaccination, cervical cancer screening, and treatment, CaCx is largely avoidable.
While CaCx screening coverage is less than 5\%, vaccination rates in Uganda are less than 25\%. It is critical to develop a system for supporting the diagnosis of cervical cancer from colposcopy images that is scientifically sound, accurate, and time-effective in order to lessen the burden of cervical cancer screening.
On the other hand, early detection of cervical lesions considerably increases the likelihood that patients will receive effective therapies. The presented deep learning model is demonstrated to be able to segment cervical precancerous lesions. We evaluate this framework at scale on a dataset of 387 annotated cervical images, and the proposed method achieves a high IoU of 0.87 and a dice coefficient of 0.92, demonstrating that the proposed system is capable of segmenting precancerous lesions with a high IoU and dice coefficient comparable to the two state-of-the-art benchmark methods.