LLMs Improve Accuracy and Completeness of Cancer Pathology Summaries | JCO Clinical Cancer Informatics

Introduction

Modern oncology care depends on rapid interpretation of increasingly complex pathology reports that integrate histopathology, immunohistochemistry and molecular profiling. Synthesizing these data into concise, clinically usable summaries is time-intensive and cognitively demanding, creating workflow burden and increasing the risk of omission in busy oncology practice.

Problem Statement

Conventional physician-authored pathology summaries are often efficient but may incompletely capture key diagnostic and genomic information, particularly as molecular testing becomes more complex and voluminous. Large language models (LLMs) offer a potential solution, but their clinical reliability, completeness and safety in summarizing oncology pathology reports require careful evaluation before integration into routine practice.

Summary

This study demonstrates that open-source LLMs can generate clinically useful summaries of complex cancer pathology reports with greater completeness than physician-authored summaries while maintaining comparable correctness. Across 94 thoracic oncology cases, most LLMs outperformed physician summaries on objective measures of fidelity and consistently captured more complete clinicopathologic and genomic information, particularly molecular findings that were frequently omitted in routine documentation. Importantly, top-performing models maintained strong factual accuracy and low rates of clinically meaningful error, suggesting that LLM-assisted summarization can reduce documentation burden without compromising clinical usability. Performance, however, was model dependent: newer systems such as DeepSeek and Llama 3.1/3.2 performed reliably, whereas older or shorter-context models were more prone to omissions, unusable outputs and clinically relevant errors. The study highlights a key practical advantage of LLMs in oncology workflows—the ability to standardize and scale synthesis of increasingly complex pathology and genomic data—while also emphasizing the need for model selection, human oversight and task-specific validation. These findings support LLM-assisted pathology summarization as a promising workflow tool to improve documentation efficiency, reduce cognitive burden and enhance clinical information accessibility in cancer care.