Computer Vision: Prediction of homologous recombination deficiency (HRD) from whole slide images (WSIs) of H&E-stained ovarian cancer biopsies via deep learning and transformers
A PofC Computer Vision project that automated pre-screening pipeline for HRD in ovarian cancer patients using commerical imaging data
Tools & Techniques
Python (pandas, matplotlib, slideflow, Pytorch, PIL, etc.), High-performance computing (HPC) platform, Bash, GitHub, etc.
Summary
Homologous recombination deficiency (HRD) is present in roughly half of patients with advanced ovarian cancer (OV), but genetic testing (GT) to determine HRD status has had limited and inequitable reach to patients in the United States, resulting in suboptimal access to PARPi treatment for those patients.
In this project, we have explored the use of computational pathology, utilizing advances in AI/Machine learning, to predict molecular status from images of the H&E-stained biopsies for OV patients. A weakly-supervised deep learning model, CLAM, was utilized to link image-derived information to HRD status and aggregate slide-level predictions without the need for manual image annotations or user-selected regions of interest. Pre-trained vision-transformers were used to generated tile-level information. A multimodal approach (attention scores combining with patient age, GIS scores, etc.) was also evaluated.
The model was trained and tested on a small commercial datasets of de-identified ovarian cancer tissue whole slide images (WSIs) and has yielded promising results: ROC AUC score of HRD+ prediction on unseen patients acheived 75%, with high sensitivity (0.8) and increased PPV (> 23%). These findings would potentially improve the prioritization of patients with a higher likelihood of HRD for GT at the pre-screening stage. Manuscript is finished and we are waiting for submission approval.
Flow Chart
- Overview of Model
Result
- Example of a case (image is obtained from public dataset. This is for illustration only). Heatmaps and tile-level predictions offers explainable and interpretable result
Notice: Data and results are under review for manuscript submission