Ensemble Deep Learning for Cervix Image Selection toward Improving Reliability in Automated Cervical Precancer Screening

Abstract

Automated Visual Examination (AVE) is a deep learning algorithm that aims to improve the effectiveness of cervical precancer screening, particularly in low- and medium-resource regions. It was trained on data from a large longitudinal study conducted by the National Cancer Institute (NCI) and has been shown to accurately identify cervices with early stages of cervical neoplasia for clinical evaluation and treatment. The algorithm processes images of the uterine cervix taken with a digital camera and alerts the user if the woman is a candidate for further evaluation. This requires that the algorithm be presented with images of the cervix, which is the object of interest, of acceptable quality, i.e., in sharp focus, with good illumination, without shadows or other occlusions, and showing the entire squamo-columnar transformation zone. Our prior work has addressed some of these constraints to help discard images that do not meet these criteria. In this work, we present a novel algorithm that determines that the image contains the cervix to a sufficient extent. Non-cervix or other inadequate images could lead to suboptimal or wrong results. Manual removal of such images is labor intensive and time-consuming, particularly in working with large retrospective collections acquired with inadequate quality control. In this work, we present a novel ensemble deep learning method to identify cervix images and non-cervix images in a smartphone-acquired cervical image dataset. The ensemble method combined the assessment of three deep learning architectures, RetinaNet, Deep SVDD, and a customized CNN (Convolutional Neural Network), each using a different strategy to arrive at its decision, i.e., object detection, one-class classification, and binary classification. We examined the performance of each individual architecture and an ensemble of all three architectures. An average accuracy and F-1 score of 91.6% and 0.890, respectively, were achieved on a separate test dataset consisting of more than 30,000 smartphone-captured images.

Keywords: cervical cancer; cervix/non-cervix; deep learning; ensemble; one-class classification.

Figure 1

Non-cervix image samples in one collected dataset.

Figure 2

Examples of cervix images used in this study, showing quality variability. 1st row: MobileODT dataset; 2nd row: Kaggle dataset; 3rd row: COCO dataset; 4th row: SEVIA dataset (only the first two images are cervix images, the rest images in this row are non-cervix image samples in this dataset).

Figure 3

RetinaNet workflow. The convolutional layers (blue block on the left) extract features in different scales (blue, yellow and green layers in the middle) followed by NMS (Non-maximum Suppression) and thresholding.

Figure 4

Deep SVDD [21] workflow from input data distribution to output data distribution. The “hypersphere” (the circle in the right subfigure) is found by the algorithm as the smallest hypersphere with center c, and radius R, that includes all (or the majority) of the target samples in the feature space.

Figure 5

Customized CNN (Convolutional Neural Network) architecture.

Figure 6

RetinaNet Results: (a) ROC from one of the test folds, and (b) and false negative error examples.

Figure 7

Validation examples of Deep SVDD approach. Left images are predicted as cervix in the validation set. Right images predicted as non-cervix images, though all of the images shown are cervix image samples with “green filter” applied.

Figure 8

Samples of cervix images that are “misclassified” as non-cervix images.