Article Text

Download PDFPDF

2022-RA-1576-ESGO Deep-learning-based endometrial segmentation and automated immune profiling from histopathological whole slide images
  1. Georgi Dzaparidze1,2,
  2. Kristi Laht1,
  3. Erik Tamp1,
  4. Heleri Taelma1 and
  5. Stella Marleen Hõlpus1
  1. 1East Tallinn Central Hospital, Tallinn, Estonia
  2. 2West Tallinn Central Hospital, Tallinn, Estonia


Introduction/Background The rapid spread of the whole slide images (WSI) combined with deep learning models can shift the qualitative approach in pathology to the quantitative one providing evenly qualitative reports worldwide. The work aimed to explore the potential computed assisted diagnostics by automated subtyping endometrial lesions with further immunohistochemical (IHC) profiling of the malignant ones, as endometrium makes a significant part of pathologists’ practice.

Methodology 721 endometrial samples for the deep learning model development and verification from the East Tallinn Central Hospital. The samples were reviewed and annotated by two pathologists and randomly divided into training and validation (271) groups; the training dataset was generated using Pathadin software. The EfficientNet-B5-based model was created as a four-class classifier (normal endometrium, hyperplasia without atypia, atypical hyperplasia, malignancy). In samples with malignancy, computer vision next detected the corresponding region on the IHC (ER, PR, p53, Her2) stained glasses and quantified it using pre-trained DeepLiif solution. DeeLiif was trained on the control samples provided with all the IHC glasses.

Results The model is the first four-type classifier for histopathological WSI classification of endometrial lesions. Tested on 271 slides from a single medical center cohort, an AUC of 0.882 was achieved, mainly failing to distinguish between atypical hyperplasia and G1 endometrioid carcinoma. For IHC, total accuracy of 0.865 was achieved, primarily failing to analyze the membranous staining of Her2.

Conclusion The algorithms successfully classified the samples and detected and analyzed the corresponding area on the IHC stained glasses, proving the concept that with proper validation and under the control of a pathologist can already be covering a part of daily routine. For further improvement, samples from different hospitals should be harvested, a model with precise diagnoses should be created, and the spatial shifting in the series of sections should be resolved.

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.