Article Text

Download PDFPDF
680 Is there a way to optimize the performance of preoperative diagnosis of patients with endometrial cancer on imaging? The artificial intelligence response
  1. Lise Lecointre1,2,3,
  2. Julia Alekseenko1,2,
  3. Matteo Pavone1,4,5,
  4. Alexandros Karargyris1,6,
  5. Jeremy Dana1,7 and
  6. Nicolas Padoy1
  1. 1Institut Hospitalo-Universitaire (IHU) Strasbourg, Institute of Image-Guided Surgery, Strasbourg, France
  2. 2ICube UMR 7357-Laboratoire des Sciences de l'Ingénieur, de l'Informatique et de l'Imagerie, CNRS, University of Strasbourg, Strasbourg, France
  3. 3Department of Gynecologic Surgery, University Hospitals of Strasbourg, Strasbourg, France
  4. 4IRCAD, Research Institute Against Digestive Cancer France, Strasbourg, Strasbourg, France
  5. 5Dipartimento per la salute della Donna e del Bambino e della Salute Pubblica, Fondazione Policlinico Universitario A. Gemelli, IRCCS, UOC Ginecologia Oncologica, Rome, Italy
  6. 6MLCommons, San Francisco, United States
  7. 7Diagnostic Radiology Department, McGill University Health Centre, Montreal, Canada


Introduction/Background Evaluation of prognostic factors is crucial in patients with endometrial cancer to ensure optimal treatment planning and accurate prognosis assessment. This study introduces an end-to-end MRI-based deep learning (DL) pipeline from tumor and uterus segmentation to deep myometrial invasion (DMI) and cervical stroma invasion (CSI) prediction to assist radiologists in pre-operative workup.

Methodology 178 pre-treatment pelvic sagittal T2-weighted images were obtained from 178 endometrial cancer patients (DMI: 90/178 – 51%, CSI: 31/178 – 17.42%). The dataset was randomly divided into a training (150 patients – 84%) and a test (28 patients – 16%) set. The proposed DL method was trained and validated with 5-fold cross-validation. This end-to-end solution (Figure 1) included a segmentation module based on a two-stage pipeline for efficient uterus segmentation and tumor localization. Tumor features were then extracted, and the prediction module trained using a vector database to offer only optimal and similar candidates for the Siamese network, which was trained in a contrastive manner.

Results The model achieved a test accuracy of 0.79 in predicting DMI and a test balanced accuracy of 0.75 in predicting CSI. In comparison, expert readers’ accuracies were 0.71 and 0.96, respectively. The accuracy rates for uterus and tumor segmentation, measured by the Dice coefficient, were 0.87 and 0.60, respectively.

Conclusion This fully automated approach achieved good-to-excellent accuracy and holds great promise in assisting radiologists with pre-operative evaluations of DMI and CSI. Moreover, it demonstrated a robust capability in accurately segmenting key regions of interest, specifically the uterus and lesions, highlighting the positive impact our solution can bring to healthcare imaging.

Disclosures None.

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.