ABSTRACT
Image quality control is a critical element in the process of data collection and cleaning. Both manual and automated analyses alike are adversely impacted by bad quality data. There are several factors that can degrade image quality and, correspondingly, there are many approaches to mitigate their negative impact. In this paper, we address image quality control toward our goal of improving the performance of automated visual evaluation (AVE) for cervical precancer screening. Specifically, we report efforts made toward classifying images into four quality categories ("unusable", "unsatisfactory", "limited", and "evaluable") and improving the quality classification performance by automatically identifying mislabeled and overly ambiguous images. The proposed new deep learning ensemble framework is an integration of several networks that consists of three main components: cervix detection, mislabel identification, and quality classification. We evaluated our method using a large dataset that comprises 87,420 images obtained from 14,183 patients through several cervical cancer studies conducted by different providers using different imaging devices in different geographic regions worldwide. The proposed ensemble approach achieved higher performance than the baseline approaches.
ABSTRACT
BACKGROUND: This cross-sectional pilot study evaluates diagnostic accuracy of live colposcopy versus static image Swede-score evaluation for detecting significant precancerous cervical lesions greater than, or equal to grade 2 severity (CIN2+). METHODS: VIA or HrHPV positive women were examined using a mobile colposcope, in a rural clinic in Kolkata, India. Live versus static Swede-score colposcopy assessments were made independently. All assessments were by gynecologists, junior or expert. Static image assessors were blinded to live scoring, patient information and final histopathology result. Primary outcome was the ability to detect CIN2+ lesions verified by directed biopsies. Diagnostic accuracy was calculated for live versus static Swede-score in detecting CIN2+ lesions, as well as for interclass correlation. RESULTS: 495 images from 94 VIA positive women were evaluated in this study. Thirteen women (13.9%) had CIN2+ on biopsy. No significant difference was found in the detection of CIN2+ lesions between live and static assessors (area under curve = 0.69 versus 0.71, p = 0.63). A Swede-score of 4+, had a sensitivity of 76.9% (95% CI 46.2-95.0%) and 84.6% (95% CI 54.6-98.1%), for live- and static-image assessment respectively. The corresponding positive predictive values were found to be 90.9% (95% CI 75.7-98.1%) and 92.6% (95% CI 75.7-99.1%). The interclass correlation was good (kappa statistic = 0.60) for the senior static assessors. CONCLUSIONS: Swede-score evaluation of static colposcopy images was found to reliably detect CIN2+ lesions in this study. Larger studies are needed to further develop the colposcopy telemedicine concept which may offer reliable guidance in management where direct specialist input is not available. TRIAL REGISTRATION: Ethical approval of the study was obtained by the Chittaranjan National Cancer Institute (CNCI) Human Research Ethics Committee (4.311/27/2014). The trial was retrospectively registered in the Clinical Trails Registry of India CTRI/2018/03/012470 .
