Ultrasonographic diagnosis of ovarian tumors through the deep convolutional neural network.

Objectives: The objective of this study was to develop and validate an ovarian tumor ultrasonographic diagnostic model based on deep convolutional neural networks (DCNN) and compare its diagnostic performance with that of human experts.

Material and methods: We collected 486 ultrasound images of 192 women with malignant ovarian tumors and 617 ultrasound images of 213 women with benign ovarian tumors, all confirmed by pathological examination. The image dataset was split into a training set and a validation set according to a 7:3 ratio. We selected 5 DCNNs to develop our model: MobileNet, Xception, Inception, ResNet and DenseNet. We compared the performance of the five models through the area under the curve (AUC), sensitivity, specificity, and accuracy. We then randomly selected 200 images from the validation set as the test set. We asked three expert radiologists to diagnose the images to compare the performance of radiologists and the DCNN model.

Results: In the validation set, AUC of DenseNet was 0.997 while AUC was 0.988 of ResNet, 0.987 of Inception, 0.968 of Xception and 0.836 of MobileNet. In the test set, the accuracy was 0.975 with the DenseNet model versus 0.825 (p < 0.0001) with the radiologists, and sensitivity was 0.975 versus 0.700 (p < 0.0001), and specificity was 0.975 versus 0.908 (p < 0.001).

Conclusions: DensNet performed better than other DCNNs and expert radiologists in identifying malignant ovarian tumors from benign ovarian tumors based on ultrasound images, a finding that needs to be further explored in clinical trials.