Abstract Background: Since this is the first study to utilize CODA (a deep-learning cell segmentation model) on prostate tumors, we explored three objectives: (1) train the CODA model to annotate morphological features of prostate tumors of different demographics to build a robust model, (2) utilize CODA to identify the spatial localization of immune cells, and (3) perform Visium and Akoya multiplexing analyses using the slides detected by CODA. African Americans (AA) are disproportionately affected by prostate cancer, as they have a higher mortality rate from the disease. Our group and others have routinely observed that tumors from African American patients present with elevated levels of immune-related gene expression. To resolve immune cell infiltration within prostate tumors, we sought to utilize CODA to quantify morphological characteristics of prostate tissue from African American, European American (EA), and Nigerian prostate cancer patients. Methods: Through successive training on each prostate cell type, the CODA model was built using samples from 6 patients who self-identified as AA and 4 patients who self-identified as EA for a total of 10 patients. All of the patients had aggressive prostate cancer with a Gleason score of 9. To build the first model, we performed 10-15 annotations for each of the nine classifications per tissue sample for 27 independent histology images. The model learned from these annotations and obtained an overall accuracy of 94% when tested on an independent annotated image. We then used the model to segment the entire cohort of 595 histology images from AA and EA patients. Interestingly, when we applied the trained CODA model to Nigerian prostate cancer patients, the overall accuracy was below 50%. Therefore, we performed that same procedure on Nigerian men with prostate tumors. After this model is optimized, we will test how the Nigerian model performs on the AA and EA images, and how the AA/EA model performs on the Nigerian images. Results: After performing a bulk quantification to calculate the tissue composition of the 10 patients, it was determined that the percent composition of cancer was similar in AA and EA samples. Prostate cancer in AA and EA were both around 7% in tissue composition (p-value=0.92). Inflammation was higher in AA than in EA men, with a % composition of 0.75 % and 0.60%, respectively (p-value=0.43). The percent composition of inflammation differed in terms of where the inflammation was located in both demographics. The percent composition of inflammation near cancer was higher in EA men than in AA men, and far from the cancer, the percent composition of inflammation was higher in AA men than in EA men. Additionally, the data collected from Visium HD and protein-phenotyping analyses will be integrated with CODA to annotate the immune cell populations at the single cell level. Conclusions: This is novel as it is the first AI CODA model on the prostate utilizing population specific specimens. A robust model successfully captured the heterogeneity of the prostate in different ethnic populations. Citation Format: Jevon Layne, Farjana Yesmin, Isra Elhussin, Ezekiel Wamble, Huixian Lin, Ezra Baraban, Ashley Kiemen, Clayton C. Yates. Training AI based CODA model for pathological analysis of prostate cancer across diverse demographics abstract. In: Proceedings of the 18th AACR Conference on the Science of Cancer Health Disparities; 2025 Sep 18-21; Baltimore, MD. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2025;34(9 Suppl):Abstract nr C014.
Layne et al. (Thu,) studied this question.