Deep Learning for Assessing Image Focus for Automated Cervical Cancer Screening

Cervical cancer is one of the leading causes of women's mortality worldwide. Early diagnosis of precancer, the highest grade of cervical intraepithelial neoplasia (CIN) prior to being called cancer, is critical in improving the survival rate. Visual Inspection with Acetic Acid (VIA) is a visual examination technique that reveals lesions with HPV infection that whiten on exposure to 5% acetic acid. Combined with HPV testing, VIA can be an effective screening technique in low resource settings. Recently, a deep learning method for screening cervical images showed the ability to detect precancer at a rate superior to human experts, particularly in women 25-49 years of age, the critical age for high yield in screening the disease. However, the method's performance depends on having good quality images. Smartphone-based enhanced cervical image assessment is a strong candidate to replace the commonly used VIA technique in rural low-resource settings but is susceptible to poor image quality - particularly out-of-focus images. Thus, detecting sharp images is a critical first step toward accurate screening for cervical precancer. We present a deep learning architecture that detects in-focus smartphone cervical images. The method was evaluated on over 4500 images acquired by a commercial cervical image acquisition framework. We examined and compared three types of deep learning networks: an object detection model (RetinaNet); fine-tuned deep learning models (VGG, Inception); and transfer learning models (VGG, Inception Feature extractor + SVM) in evaluating the sharpness of the images. The highest technical image quality assessment accuracy we obtained was 94%.