Presentation Authors: Hao G. Nguyen, Bogdana Schmidt*, Ehsan Hosseini-Asl, Clarence So, Richard Socher, Caiming Xiong, Lingru Xue, Peter R Carroll, Matthew R. Cooperberg, San Francisco, CA
Introduction: IF performed on tissue microarrays (TMA) is a proven platform for both rapid and cost-effective screening and validation of biomarkers but has been limited to the arduous, sometime subjective interpretation of the visual assessment through an IF microscope. We implemented an AI model to automate the analysis of biomarkers by recognizing specific expression patterns of the markers of interest (Ki 67, Erg, PTEN, c-MYC, AR) in epithelial cells and normal stromal tissue to translate the finding into predictions of recurrence and metastasis after radical prostatectomy.
Methods: A TMA was constructed consisting of 648 samples (424 tumor and 224 normal tissue) generated from patients who underwent radical prostatectomy for localized prostate cancer at UCSF and had been previously subjected to RNA-based biomarker assessment. The TMA was subjected to IF staining using antibodies against Ki 67, ERG, PTEN, c-MYC, AR and CK8 and analyzed for differential expression using standardized microscopy and an AI model. Relative mean IF intensity was used to extrapolate differential expression in normal tissue and cancerous tissue. AI model was designed to recognize both patterns and details at the pixel level, by discriminating epithelium, stroma, and artifacts, using a training cohort. The trained model was then validated using a separate cohort from the TMA. Predicted data from the deep learning model were then compared to the manual IF analysis.
Results: The analysis using Ki-67 staining and ERG positivity and expression level generated by the AI algorithm showed only a 5% variance from AI algorithm vs manual ascertainment. The model was able to pick out ERG positive tumors with 100% accuracy. AI algorithm maintained accuracy despite images and data variance from artifacts. Furthermore, the AI model has the ability to improve accuracy after each round of modification and feedback back through training cohort.
Conclusions: We demonstrated that our new AI model produces similar outcomes with high accuracy and robustness as manual quantification but with more efficiency, cost effectiveness, and objectivity.
Source of Funding: Department of Defense grant W81XWH-15-1-0460.