Development of image preprocessing toolbox to boost cervical cancer detection

Globally, inequitable resources and access cause cervical cancer remain the fourth leading cause of death in women. Risk stratification with human papilloma virus (HPV) screening is fundamental in identifying vulnerable populations. Visual inspection with acetic acid (VIA) however, is often required to triage HPV+ individuals. While triage with VIA can allow for resource prioritization, it can be highly subjective due to variations in provider practice. Deep learning diagnostic algorithms have emerged as assistive tools to providers. Though major strides have been made, datasets informing these algorithms can be impacted by image quality limitations such as blur and transformation zone visibility. In this work we present the use of an image quality control toolbox using deep and machine learning to outline biological and nonbiological image quality factors. Deep learning techniques are employed sequentially to identify blurry images and detect the presence of columnar epithelium (CE), indicating transformation zone visibility. Automated ground truth annotation and a YOLOv5 object detection model localize a cervix region of interest and detect blur within a multi-device image database. On this refined set, a semi-supervised U-Net model is trained to segment CE tissue, marking the inner boundary of the key squamocolumnar junction. The blur detection model achieved a mean average precision score of 0.9 while yielding an accuracy of 0.72 when compared to expert annotation. Initial model results indicate successful CE segmentation, achieving a validation accuracy of 0.92. These tools build an image processing toolbox for construction of informative datasets and retroactive evaluation developing diagnostic algorithms.