Introduction:

Immunofluorescence (IF) performed on tissue microarrays (TMA) is a proven platform for both rapid and cost-effective screening and validation of biomarkers but is limited by thearduous and subjective human visual assessment with an IF microscope. We aim to implement deep learning-based artificial intelligence (AI) models to automate and speed up the analysis of numerous biomarkers such as Ki 67, Erg, PTEN, c-MYC, AR (androgen receptor), by using various algorithms to recognize specific patterns of expression in epithelial cells and normal stromal tissue for each marker of interest in order to translate the findings into prediction models of recurrence and metastasis after surgery.

Methods:

A TMA was constructed consisting of 648 samples (424 tumor, 224 normal tissue) generated from radical prostatectomy specimens done for localized prostate cancer. Each had been previously subjected to RNA-based biomarker assessment. IF staining was performed on the TMA using antibodies against Ki 67, ERG, PTEN, c-MYC, AR and CK8 (cytoskeleton 8) and analyzed for differential expression using “gold standard” standardized manual microscopy and using a deep learning-based algorithm. Analysis was done blinded to any clinicopathological data.

For the manual microscopy, relative mean fluorescence intensity was used to extrapolate the differential expression in normal adjacent tissue to that of cancerous tissue. Then, AI algorithms were designed to recognize both broad patterns and specific details of the digitized images at pixel level, by discriminating epithelium, stroma, and artifacts, using a training cohort. During its development, the model learns to accurately disentangle overlapping regions and touching cells, by leveraging prior information of cell structure computed by pixel-intensity based algorithms. To do so, the Otsu method thresholding algorithm combined with mean shift clustering was employed, to find the cell centers, followed by a level-set algorithm, to compute the initial cell boundaries. These predictions were then combined with pixel predictions of a fully-convolutional deep model to refine the regions of three overlapping tissues, i.e. epithelium, stroma, and artifacts (Figure 1). DAPI was used for nuclear staining. The trained model was then validated using a separate cohort from the TMA. Predicted data from the algorithm were then compared to the data from the manual microscopy.

Results:

We started with Ki-67 and ERG stainings of the constructed TMA. The analysis using Ki-67 and ERG positivity and expression levels generated by the AI algorithm showed only a 5% variance compared to the manually generated “gold standard” results.  The AI algorithm was able to pick out which tumor were positive for ERG with 100% accuracy, meaning accuracy was maintained despite data variance from artifacts. Furthermore, the AI program had the ability to improve its accuracy after each iteration of modifications and feedback through the training cohort.

Figure 1: ERG expression pipeline showing original staining of ERG expression (red), artifact (green), DAPI nuclear staining (blue) and decomposed image in black and white and final composite with quantification from the AI algorithm.

Conclusion:

We demonstrated that our new AI algorithm produces similar outcomes with high accuracy and robustness when compared to manual quantification but with more efficiency, cost effectiveness, and objectivity. We are now developing more complex algorithms that will include the differential pattern of expression of PTEN, MYC and AR in regions high grade cancer, as well as CK8 (to help better distinguish epithelial cells from stromal cells), with the objectives of streamlining discovery and validation of novel biomarkers for lethal prostate cancer.

Funding: N/A