Developing a deep learning algorithm for automated p53 immunohistochemistry digital image analysis in prostate cancer.

223 Background: p53 immunohistochemistry (IHC) is a prognostic marker in prostate cancer (PCa), but manual scoring by pathologists is time-consuming and prone to variation. Digital image analysis (DIA) offers a potential solution to enhance speed, accuracy, and reproducibility. This study aims to develop a DIA algorithm for scoring p53 IHC in PCa and compare its performance with manual scoring. Methods: Forty patients with adequate archival PCa tissue were randomly selected, and representative sections were stained for p53 IHC. Two pathologists independently scored p53 nuclear expression based on intensity (0, 1+, 2+, 3+) and percentage of expression in tumor cells (0%, < 1%, 1–5%, > 5%). Whole slide images (WSIs) were analyzed using a Visiopharm batch processing workflow. In a training set of 14 samples, each WSI was manually segmented into benign, tumor, and background regions. The algorithm with the highest accuracy for tumor detection was selected for p53 scoring. Artifacts generating false positives, such as surgical ink and pigments, were filtered by an Artifact Detection APP. Problematic segmentation issues, including hemosiderin-laden macrophages and folded tissue sections, were manually excluded. A Tumor Detection APP identified tumor regions of interest. A Nuclei Detection APP identified the tumor nuclei and was calibrated to differentiate 3,3′-diaminobenzidine (DAB) intensity thresholds for scores of 0–3+ on the evaluation set of 40 samples. The final application was applied to the evaluation set, and results were compared with manual pathologist scores. All image analysis algorithms were trained using convolutional neural networks (CNNs) on pathologist-guided training sets within the Visiopharm platform, executed as part of the whole slide analysis protocol. Results: Comparison of automated DIA p53 scores with manual consensus scores (Table) revealed a Kendall’s coefficient of concordance of 0.81 (p = 0.008), indicating strong agreement. The Goodman-Kruskal Gamma coefficient showed a high correlation of 0.8 (95% CI: 0.6-1.0). The estimated Volume Under the ROC Surface (VUS) for the 4-level ordinal score was 0.68 (SE = 0.11), suggesting moderate to high discriminatory power in distinguishing among the four scoring levels based on the percentage of positive nuclei. Conclusions: The strong inter-rater agreement and high correlation between manual and DIA scores suggest that the deep learning algorithm assesses p53 IHC in PCa with moderate to high discriminatory power. Robust slide selection is crucial to minimize false positives. These findings highlight the potential of automated DIA to enhance efficiency and standardization of p53 IHC scoring. Distribution of manual p53 scores (rows) and automated digital image analysis (DIA) p53 scores (columns) by intensity. Manual p53 scores (n) Automated DIA p53 scores (n) 0 1+ 2+ 3+ 0 18 5 1 2 1+ 1 5 3 1 2+ 0 0 2 0 3+ 0 0 0 2