Comparison of different focus measuring functions in low-cost cervical imaging

Approximately 90% of deaths from cervical cancer deaths occur in low and middle income countries. Imaging the cervix is necessary for diagnosis, but the imaging geometry of the cervix (from the outside, through a speculum that serves as an aperture stop) is non-trivial. Blur is present in some form in most cervical images. The unique imaging conditions necessitate assessing image sharpness in real time, yet it is unclear which focus assessment method is optimal. In this paper, 31 different focus measuring functions were analyzed grouped into the following categories: gradient, Laplacian, statistical, wavelet, and miscellaneous. A set of 3,458 cervical images from the OPTICS (Optimizing Prevention by Taking Images for Cervical cancer Screening (and triage)) trial were annotated for sharpness and the location (bounding box) of the cervix by a colposcopy expert. There were 3 classes of sharpness: blurry (0), slightly blurry/sharp (1), and sharp (2). The area within the bounding box was evaluated for sharpness using the focus measuring functions and compared to the manual annotations. The area under the receiver operating characteristic (ROC) curve (AUC) was calculated. The ROC AUC analysis showed that gradient-based methods performed best under both the ‘sharp versus not sharp’ and ‘blurry versus not blurry’ scenarios, with ROC AUC values reaching values close to 0.8. Additionally, the runtime of each function was evaluated on the Samsung S20 phone and the Samsung J530 phone, showing a large variation across the different focus measuring functions.