Article Text

Download PDFPDF
1291 Performance of an artificial intelligence model applied to transvaginal ultrasound images in detecting endometrial cancer and endometrial hyperplasia with atypia in patients with postmenopausal bleeding
  1. Diletta Fumagalli1,
  2. Emilia Palmieri1,2,
  3. Sana Khan3,
  4. Adriana Gregory3,
  5. Luigi ADe Vitis1,4,
  6. Ilaria Capasso1,2,
  7. Tommaso Occhiali1,
  8. Hiroaki Takahashi3,
  9. Bohyun Kim3,
  10. Abimbola O Famuyide1,
  11. Daniel M Breitkopf1,
  12. Carrie L Langstraat1,
  13. Evelyn A Reynolds1,
  14. Christopher C Destephano5,
  15. Kristina A Butler6,
  16. Angela J Fought7,
  17. Michaela E Mcgree7,
  18. Andrea Mariani1,
  19. Gretchen E Glaser1 and
  20. Timothy L Kline3
  1. 1Department of Obstetrics and Gynecology, Mayo Clinic, Rochester, USA
  2. 2Department of Women, Children and Public Health Sciences, Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Rome, Italy
  3. 3Department of Radiology, Mayo Clinic, Rochester, USA
  4. 4Department of Gynecological Surgery, European Institute of Oncology, Milan, Italy
  5. 5Department of Obstetrics and Gynecology, Jacksonville, USA
  6. 6Department of Obstetrics and Gynecology, Mayo Clinic, Phoenix, USA
  7. 7Department of Quantitative Health Sciences, Mayo Clinic, Rochester, USA


Introduction/Background Transvaginal ultrasound (TVUS) is a method used to triage patients experiencing postmenopausal bleeding (PMB). TVUS detects endometrial atypical hyperplasia (EAH) and endometrial cancer (EC). We aimed to develop and evaluate an Artificial Intelligence (AI) pipeline on TVUS images of PMB patients to automatically segment endometrial images and classify them as EAH/EC or benign.

Methodology Our study included 480 patients with PMB who underwent TVUS and endometrial biopsy at Mayo Clinic (01/2016–09/2023). For each patient, two static TVUS images were manually annotated by radiologists. Three different labels were used to segment: (i)endometrium, (ii)endometrial cystic changes, and (iii)non-endometrial fluid (pyometra/hematometra). First, a model for automated segmentation of endometrial structures was implemented using a deep learning convolutional neural network based on the U-Net architecture. A cohort of 287 patients was used to train with 5-fold cross-validation and an ensemble model was built from these five models. Second, a deep learning algorithm for the endometrial classification (EAH/EC or benign) with 3-fold cross-validation was trained and evaluated. In the development of the second model, a range of architectural frameworks were investigated (VGG16, MobileNetV2, and ResNet), optimizing various hyperparameters (e.g., patch size, number of patches, batch size). The 193 remaining patients constituted an independent test set for model evaluation.

Results Regarding patient characteristics, no significant differences between the cohorts were observed (Table 1). The deep learning-based automated segmentation when compared to manual segmentations showed good agreement. The mean±standard deviation of the Dice similarity coefficient was 0.71±0.26 for endometrium. The current deep learning algorithms for the classification model achieved an area under the receiver operating characteristic curve in the range of 0.60–0.65 on the validation set.

Conclusion We developed and evaluated a segmentation-and-classification AI pipeline to detect EAH/EC from TVUS images in PMB women. Manual and automatic segmentation showed good agreement. Further testing and validation on broader datasets are anticipated in 2024.

Disclosures No disclosures.

Abstract 1291 Table 1

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.