Article Text

Download PDFPDF

1009 Deep learning for improved detection of premalignant lesions in the fallopian tube, a proof of concept
  1. J Bogaerts1,
  2. M Van Bommel2,
  3. J Linmans1,
  4. N Van den Hork1,
  5. J Bulten1,
  6. J De Hullu2,
  7. M Simons1 and
  8. J Van der Laak1
  1. 1Radboud University Medical Center, Pathology, Nijmegen, Netherlands
  2. 2Radboud University Medical Center, Gynecologic oncology, Nijmegen, Netherlands


Introduction/Background*Risk reducing salpingo-oophorectomy is an effective intervention to reduce the risk of high grade serous carcinoma (HGSC) in patients with a BRCA1/2 pathogenic variant (PV), but results in significant short and long-term health risks. New interventions, such as risk-reducing salpingectomy with delayed oophorectomy are promising. In this alternative approach, the detection of serous tubal intraepithelial carcinoma (STIC) as precursor to HGSC, has become more important. The detection of STIC indicates an increased risk for HGSC and would prompt for an immediate oophorectomy. Unfortunately, reproducibility of STIC diagnosis is only moderate, even among experienced gynecological pathologists. The aim of this pilot study is to develop and validate an AI algorithm for automated detection of potential STIC lesions in scanned H&E slides, to aid the pathologist in diagnosing STIC

Methodology We collected and digitalised 60 cases of STIC and 65 control cases. STIC diagnosis was confirmed using p53 and Ki-67 immunohistochemical stainings (IHC). The dataset was split into 50 cases for training, five for validation and five for testing. We developed a Convolutional Neural Network and compared two appraoches: directly detecting STIC (one-step) or first detecting all epithelium and subsequently detecting STIC within epithelial regions (two-step). Additionaly, we evaluated whether we could improve the network by enriching the training data with hard negative examples.

Result(s)*We found that the optimal configuration for detection of STIC was the two-step approach, with training set enrichment by hard negatives. This network reached an area under the receiver operating curve of 0.90 (figure 1). Visual inspection of cases in the test set showed concordance between the model output, p53 and Ki67 IHC, and pathologists‘ annotations (see example model output on H&E stained tissue in figure 2.

Conclusion*We present a convolutional neural network that can succesfully detect STIC lesions in whole slide images. AI has the potential to aid the pathologist in the detection of STIC and assist in producing more accurate and consistent diagnosis. Additional performance and robustness is expected to be achieved by expansion of the dataset.

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.