Beyond Digital Pathology: Leveraging AI to Improve Cancer Detection and LLMs to Personalize Care

Background: In our department we successfully implemented digital pathology for primary diagnosis in Sept 2021 and have been 100% digital since July 2023. The all-digital workflow has been embraced by all faculty, residents, fellows, and technical and administrative staff, making our transformation one of the first and most complete in the US. This provided a firm foundation for incorporating powerful AI tools to realize the full promise of Digital Pathology.Methods: Over the past year we validated and deployed commercially available Computer Vision AI assistance modules (from Ibex Analytics) for cancer detection, now in routine use with all prostate and breast biopsies. Using archival cases, we evaluated each module's performance relative to the ground truth original final diagnosis and calculated standard statistical metrics for clinical validation. We also developed a Customized Large Language Model (Custom LLM) using detailed prompt engineering and clinical logic grounded in evidence-based national guidelines such as those from the National Comprehensive Cancer Network (NCCN).Results: In-house validation of the prostate cancer module yielded a sensitivity of 1.0, specificity of 0.99, and AUC of 0.995. Corresponding values for breast cancer detection were 0.98, 0.99, and 0.985. Post-deployment, we continue to observe high accuracy and reliability. Additional data show the utility of incorporating these AI modules in routine practice, with positive impacts on turnaround time, pathologist productivity, clinician satisfaction, and trainee education.Initial validations show that our Custom LLM for Prostate Cancer performs reliable and accurate extraction of clinically important information from uploaded structured and unstructured data, including clinical notes, lab values, radiology findings, and pathology reports, to generate integrative diagnostic reports with personalized risk stratification and management recommendations, citing applicable guideline statements. It can also generate an optional patient-friendly report in clear, non-technical language to support informed discussions with clinicians.The Custom LLM has received strong endorsement from Radiology and Urology colleagues, who are collaborating with us to validate and safely deploy it within an EMR-integrated environment with built-in privacy and safety guardrails, including PHI and protected class filters. Additional Custom LLMs (breast, colon) are being developed using the same general approach.Conclusions:Digital Pathology is not an end unto itself. Rather, it is the foundation that enables departments to benefit from advances in AI-through computer vision tools for diagnostic assistance or generative AI for integrated, personalized reports. These innovations help bridge clinical silos, support shared decision-making, and deliver value across the health system.

Learning Objectives:

1. Describe the process by which commercial AI tools can be validated and incorporated into routine pathology practice
2. Explain the development of a Custom Large Language Model, starting with a commercially available Large Language Foundation Model such as Chat GPT
3. Describe several ways in which Digital Pathology and AI can improve quality and enhance the overall value that Pathologists can deliver to clinicians and patients