Multicontrast Pocket Colposcopy Cervical Cancer Diagnostic Algorithm for Referral Populations

Abstract

Objective and Impact Statement. We use deep learning models to classify cervix images-collected with a low-cost, portable Pocket colposcope-with biopsy-confirmed high-grade precancer and cancer. We boost classification performance on a screened-positive population by using a class-balanced loss and incorporating green-light colposcopy image pairs, which come at no additional cost to the provider. Introduction. Because the majority of the 300,000 annual deaths due to cervical cancer occur in countries with low- or middle-Human Development Indices, an automated classification algorithm could overcome limitations caused by the low prevalence of trained professionals and diagnostic variability in provider visual interpretations. Methods. Our dataset consists of cervical images ($n=\mathrm{1,760}$) from 880 patient visits. After optimizing the network architecture and incorporating a weighted loss function, we explore two methods of incorporating green light image pairs into the network to boost the classification performance and sensitivity of our model on a test set. Results. We achieve an area under the receiver-operator characteristic curve, sensitivity, and specificity of 0.87, 75%, and 88%, respectively. The addition of the class-balanced loss and green light cervical contrast to a Resnet-18 backbone results in a 2.5 times improvement in sensitivity. Conclusion. Our methodology, which has already been tested on a prescreened population, can boost classification performance and, in the future, be coupled with Pap smear or HPV triaging, thereby broadening access to early detection of precursor lesions before they advance to cancer.

Figure 1

Dataset summary. The Pocket colposcopy dataset contained acetic acid images from 1,281 unique patient visits. Of those, 101 were excluded due to the absence of completed diagnostic results—either an indication of negative upon visual inspection or a biopsy or LEEP result. In total, 880 visits contained acetic acid and green light colposcopy pairs with their associated diagnostic result, which were used to train and test our model (a). Representative Pocket colposcope images using acetic acid and green light contrasts on cervixes diagnosed as normal under visual inspection with a colposcope from a screened Pap+ population (b, c), biopsy-confirmed low-grade precancer from a screened Pap+ population (d, e), biopsy-confirmed high-grade precancer from a screened Pap+ population (f, g), and biopsy-confirmed cancer from a screened VIA+ population (h, i).

Figure 2

Representative YOLOv3 cervix detection results show removal of extraneous content from vaginal walls and speculum. YOLOv3 is used for automated cervix detection after training on images with a hand-drawn bounding box surrounding the cervix area. After automatic bounding box detection, the cervix is cropped and reshaped for entry into the classification network. Automated detection is necessary for the removal of extraneous content from vaginal walls (a–c) and the speculum pointed out with the black arrows (d–f). Both representative images are diagnosed as high-grade precancer and were captured from a Pap+ screened population.

Figure 3

Loss function selection. Comparisons of test set performance metrics of the Resnet-18 model without and with the weighted loss function, showing an improved sensitivity and less variation when the model was trained on 756 AA images and tested on 124 AA images. Sensitivity and specificity values, were calculated using a CNN output threshold of 0.5.

Figure 4

The combination of acetic acid and green light images improves overall classification performance compared to a single source of contrast for the parallel combination method. When training the network on image stacks of both acetic acid and green light images, the AUC does not show an improvement compared to acetic acid or green light alone (a). When features for both acetic acid and green light are extracted in parallel and then combined before the network’s fully connected layer, however, the AUC improves to 0.87 with the combination of green light and acetic acid images (b). In addition, there is a large performance improvement in sensitivity, increasing from 0.30 with acetic acid only to 0.75 when the combination of two contrast sources is used.

Figure 5

Key steps of the analysis process. Initially, hand-drawn bounding boxes were used to label cervix locations to train a YOLOv3 cervix detector to automatically remove extraneous content from the colposcopy image (a). Next, the trained cervix detector automatically generated the bounding boxes for the training images of the classification algorithm. The bounding boxes were used to crop and resize the cervix images preceding the removal of specular reflection. The preprocessed image and the associated label were then used to train the classification model, which was tested with and without a weighted loss function (b).

Figure 6

Dual-contrast feature extraction method. Using the Resnet-18 CNN architecture, we explored two approaches for combining acetic acid and green light colposcopy images taken from a single patient. The “image stack” method concatenates the two RGB images to create a 6-channel input into the neural network (a). The “parallel feature extraction method” tunes two networks, one for each contrast source, to extract features independently before combining the fully connected layers for classification (b).