BMC Cancer (May 2022)

Leveraging artificial intelligence to predict ERG gene fusion status in prostate cancer

  • Vipulkumar Dadhania,
  • Daniel Gonzalez,
  • Mustafa Yousif,
  • Jerome Cheng,
  • Todd M. Morgan,
  • Daniel E. Spratt,
  • Zachery R. Reichert,
  • Rahul Mannan,
  • Xiaoming Wang,
  • Anya Chinnaiyan,
  • Xuhong Cao,
  • Saravana M. Dhanasekaran,
  • Arul M. Chinnaiyan,
  • Liron Pantanowitz,
  • Rohit Mehra

DOI
https://doi.org/10.1186/s12885-022-09559-4
Journal volume & issue
Vol. 22, no. 1
pp. 1 – 9

Abstract

Read online

Abstract Background TMPRSS2-ERG gene rearrangement, the most common E26 transformation specific (ETS) gene fusion within prostate cancer, is known to contribute to the pathogenesis of this disease and carries diagnostic annotations for prostate cancer patients clinically. The ERG rearrangement status in prostatic adenocarcinoma currently cannot be reliably identified from histologic features on H&E-stained slides alone and hence requires ancillary studies such as immunohistochemistry (IHC), fluorescent in situ hybridization (FISH) or next generation sequencing (NGS) for identification. Methods Objective We accordingly sought to develop a deep learning-based algorithm to identify ERG rearrangement status in prostatic adenocarcinoma based on digitized slides of H&E morphology alone. Design Setting, and Participants: Whole slide images from 392 in-house and TCGA cases were employed and annotated using QuPath. Image patches of 224 × 224 pixel were exported at 10 ×, 20 ×, and 40 × for input into a deep learning model based on MobileNetV2 convolutional neural network architecture pre-trained on ImageNet. A separate model was trained for each magnification. Training and test datasets consisted of 261 cases and 131 cases, respectively. The output of the model included a prediction of ERG-positive (ERG rearranged) or ERG-negative (ERG not rearranged) status for each input patch. Outcome measurements and statistical analysis: Various accuracy measurements including area under the curve (AUC) of the receiver operating characteristic (ROC) curves were used to evaluate the deep learning model. Results and Limitations All models showed similar ROC curves with AUC results ranging between 0.82 and 0.85. The sensitivity and specificity of these models were 75.0% and 83.1% (20 × model), respectively. Conclusions A deep learning-based model can successfully predict ERG rearrangement status in the majority of prostatic adenocarcinomas utilizing only H&E-stained digital slides. Such an artificial intelligence-based model can eliminate the need for using extra tumor tissue to perform ancillary studies in order to assess for ERG gene rearrangement in prostatic adenocarcinoma.

Keywords