AI-driven screening of equivocal prostate IHC cases: development and validation of a screening and cancer detection framework

Artificial intelligence has been proposed as a solution to meet increasing demand for the diagnostic services of pathologists. Prostate biopsies are a significant source of this demand, and substantial fraction of these biopsies require immunohistochemistry staining, which adds work, time and cost to the diagnostic process. Equivocal cases, which often prompt the use of immunohistochemistry (IHC), present the greatest challenge to artificial intelligence cancer detection tools, but also represent the area where properly trained systems could offer the most clinical value by reducing the need for ancillary testing for cancer confirmation. From August 2021 to April 2023, we scanned 25,570 slides from 1,641 patients. We investigated the performance of institutionally developed models for prostate cancer detection using digital pathology images, with an aim for reducing work, turnaround time, and costs related to the ordering of immunohistochemistry equivocal cases. We advanced complementary sensitive and specific models to screen these challenging cases, aiming to identify slides that could be confidently diagnosed without any ancillary immunohistochemistry tests. Additionally, we compared the performance of a prostate-specific model to a general-purpose foundation model for screening and cancer detection. Our screening models correctly classified 55% of challenging equivocal blocks where immunohistochemistry was ordered with a 1.4% error rate. We found that the foundation model achieved higher screening rates (i.e., percentage of cases where IHC could be avoided), but this came at the cost of uniformly higher error rates and significantly greater computational demands. When trained as a standalone prostate cancer detection system, our model demonstrated high concordance with pathologist ground truth, achieving an area under the curve of 98.5%, sensitivity of 95.0%, and specificity of 97.8%. Computational models can aid in the diagnosis of prostate cancer and can effectively screen challenging prostate biopsy cases, reducing unnecessary IHC utilization and helping to optimize pathology workflows.