Approaches to blur reduction in cervical images for automated visual evaluation
Reuven Weiser, Francisca Morgado, Pedro Dias, Kelwin Fernandes, David Levitz · 2023
Background: Cervical cancer is a significant burden in many health systems in low and middle income countries (LMICs). Recently, automated visual evaluation (AVE) – using artificial intelligence to analyze cervical images at the point of care (PoC) – has been gaining interest as a new diagnostic test in LMICs. Multiple studies showed that blur (defocus) is the most common challenge to capturing cervical images that are adequate for evaluation by AVE. Methods to reduce blur in cervical images are critical, yet auto-focus functionality degrades when placing an auxiliary lens on a phone. Methods: A cervical image quality analysis algorithm that included blur assessment was developed into an Android application. This algorithm includes an auto-focus module, and secondary blur assessment using deep learning (DL). The auto-focus module was evaluated by bench testing on static cervix images. Two DL approaches (supervised and self-supervised models) were compared against an external dataset. Results and Discussion: A frame by frame analysis on the Samsung J530 and A52, each imaging 3 static images, verified the frame with the least blurry image was selected. The average time for one auto-focus sweep was 8367 ± 630 ms and 7555 ± 146 ms for the J530 and A52, respectively. Within the obstructions detector, the self-supervised model performed better under high blur, with area under the receiver operating characteristic (ROC) curve (AUC) as high as 0.888, while the supervised model performed better with less blur, with ROC AUC values reaching 0.735. To our knowledge, this is the first working targeted auto-focus for cervical imaging.