Classification of cervical neoplasms on colposcopic photography using deep learning

Abstract

Colposcopy is widely used to detect cervical cancers, but experienced physicians who are needed for an accurate diagnosis are lacking in developing countries. Artificial intelligence (AI) has been recently used in computer-aided diagnosis showing remarkable promise. In this study, we developed and validated deep learning models to automatically classify cervical neoplasms on colposcopic photographs. Pre-trained convolutional neural networks were fine-tuned for two grading systems: the cervical intraepithelial neoplasia (CIN) system and the lower anogenital squamous terminology (LAST) system. The multi-class classification accuracies of the networks for the CIN system in the test dataset were 48.6 ± 1.3% by Inception-Resnet-v2 and 51.7 ± 5.2% by Resnet-152. The accuracies for the LAST system were 71.8 ± 1.8% and 74.7 ± 1.8%, respectively. The area under the curve (AUC) for discriminating high-risk lesions from low-risk lesions by Resnet-152 was 0.781 ± 0.020 for the CIN system and 0.708 ± 0.024 for the LAST system. The lesions requiring biopsy were also detected efficiently (AUC, 0.947 ± 0.030 by Resnet-152), and presented meaningfully on attention maps. These results may indicate the potential of the application of AI for automated reading of colposcopic photographs.

Figure 1

Participant enrollment diagram (A) and classification systems of cervical lesions (B).

Figure 2

Representative examples of an original image (A) and the augmented images for cervical neoplastic lesion: a horizontally-flipped image (B), vertically-flipped image (C), and horizontally and vertically-flipped image (D).

Figure 3

Heatmap of the confusion matrix of the multiclass classification of cervical lesions on colposcopic photographs by the best-performing Resnet-152 model. (A) the CIN system (B) the LAST system. The figure was created using Python version 3.6.8, sklearn library version 0.21.2 and matplotlip library version 3.1.0.

Figure 4

Receiver operating characteristic curves of the best-performing Resnet-152 models for the binary classification of the CIN and LAST system, and for determining the need to biopsy.

Figure 5

Class activation map for the classification of high-risk and low-risk cervical lesions on colposcopic photographs using a convolutional neural network based on (A) the CIN system or (B) the LAST system.