Pelvic ultrasound-based deep learning models for accurate diagnosis of ovarian cancer: Retrospective multicenter study

초록

Background: Although pelvic ultrasound is useful modality to diagnose ovarian cancer, there is few studies to evaluate the performance of deep learning model. Development of deep learning based on aggregating ultrasound images and clinical information such as age and CA-125 data might be beneficial to differentiated malignancy from benign ovarian cyst. The objective of this study is to build a deep learning model with advanced accuracy of differential diagnosis between malignant and benign lesions of ovary. Methods: Pelvic ultrasound images, information of age and ca-125 level at diagnosis from patients diagnosed with ovarian cancer or benign cyst during 2015-2022 were retrieved. The images were segmented into cystic and solid component and processed data for feature extraction. For convolutional neural network model development, patients were assigned to training dataset (565 patients), validation dataset (76 patients), and test dataset (163 patients). After feature extraction using this convolutional neural network, we also used age and tumor marker to classify these images as either malignant or benign. Using these datasets, we assessed the diagnostic value of the deep learning model. A new AI model to utilize image feature and clinical information together was developed. This model separates feature extractor and classifier from the architecture of conventional AI models, and the classifier receives clinical information data in addition to image information. ResNet50 and DenseNet121 were used in the conventional AI model. Results: In total, 804 patients who had ovarian tumors were included; 446 benign and 358 malignant. When pelvic ultrasound image of ovary was only used, ResNet50 showed AUC of 0.84 and DenseNet121 had AUC of 0.82 for this model to detect ovarian cancer. However, when segmented solid portion image and clinical data was combined with the ovary image, ResNet50 showed AUC of 0.95 and DenseNet121 had AUC of 0.96 for this model to detect ovarian cancer. Using the test data set, ResNet50 and DenseNet121 had sensitivities of 90% and 81%, specificity of 93% and 97%, positive predictive value of 92% and 95%, and negative predictive value of 92% and 86% to detect ovarian cancer, respectively. Conclusions: Binary classification model based on deep learning algorithms using pelvic ultrasound images can distinguish between benign and malignant ovarian tumors accurately. Segmentation of solid portion and clinical information of age and CA-125 at diagnosis combined with the pelvic ultrasound images increased the accuracy of the classification model.