Deep learning-enabled pelvic ultrasound images for accurate diagnosis of ovarian cancer in China : a retrospective, multicentre, diagnostic study

Research output: Journal Publications and Reviews (RGC: 21, 22, 62)21_Publication in refereed journalpeer-review

View graph of relations

Author(s)

  • 88 authors, including
  • Lingxi Chen

Detail(s)

Original languageEnglish
Pages (from-to)e179-e187
Journal / PublicationThe Lancet Digital Health
Volume4
Issue number3
Publication statusPublished - 1 Mar 2022
Externally publishedYes

Link(s)

Abstract

Background: Ultrasound is a critical non-invasive test for preoperative diagnosis of ovarian cancer. Deep learning is making advances in image-recognition tasks; therefore, we aimed to develop a deep convolutional neural network (DCNN) model that automates evaluation of ultrasound images and to facilitate a more accurate diagnosis of ovarian cancer than existing methods. 
Methods: In this retrospective, multicentre, diagnostic study, we collected pelvic ultrasound images from ten hospitals across China between September 2003, and May 2019. We included consecutive adult patients (aged ≥18 years) with adnexal lesions in ultrasonography and healthy controls and excluded duplicated cases and patients without adnexa or pathological diagnosis. For DCNN model development, patients were assigned to the training dataset (34 488 images of 3755 patients with ovarian cancer, 541 442 images of 101 777 controls). For model validation, patients were assigned to the internal validation dataset (3031 images of 266 patients with ovarian cancer, 5385 images of 602 with benign adnexal lesions), external validation datasets 1 (486 images of 67 with ovarian cancer, 933 images of 268 with benign adnexal lesions), and 2 (1253 images of 166 with ovarian cancer, 5257 images of 723 benign adnexal lesions). Using these datasets, we assessed the diagnostic value of DCNN, compared DCNN with 35 radiologists, and explored whether DCNN could augment the diagnostic accuracy of six radiologists. Pathological diagnosis was the reference standard. 
Findings: For DCNN to detect ovarian cancer, AUC was 0·911 (95% CI 0·886–0·936) in the internal dataset, 0·870 (95% CI 0·822–0·918) in external validation dataset 1, and 0·831 (95% CI 0·793–0·869) in external validation dataset 2. The DCNN model was more accurate than radiologists at detecting ovarian cancer in the internal dataset (88·8% vs 85·7%) and external validation dataset 1 (86·9% vs 81·1%). Accuracy and sensitivity of diagnosis increased more after DCNN-assisted diagnosis than assessment by radiologists alone (87·6% [85·0–90·2] vs 78·3% [72·1–84·5], p<0·0001; 82·7% [78·5–86·9] vs 70·4% [59·1–81·7], p<0·0001). The average accuracy of DCNN-assisted evaluations for six radiologists reached 0·876 and were significantly augmented when they were DCNN-assisted (p<0·05). 
Interpretation: The performance of DCNN-enabled ultrasound exceeded the average diagnostic level of radiologists matched the level of expert ultrasound image readers, and augmented radiologists’ accuracy. However, these observations warrant further investigations in prospective studies or randomised clinical trials. Funding: National Key Basic Research Program of China, National Sci-Tech Support Projects, and National Natural Science Foundation of China.

Research Area(s)

Citation Format(s)

Download Statistics

No data available