Artificial intelligence models can outperform physicians in summarizing complex cancer pathology reports, according to new research from Northwestern Medicine. The study highlights the growing potential of AI to support clinicians in managing increasingly data-intensive oncology workflows.
Advances in biomarker testing and longer patient survival are driving a surge in the volume and complexity of pathology reports. These documents often include longitudinal data from multiple institutions, requiring clinicians to rapidly synthesize detailed histopathological, immunohistochemical and genomic information.
Against this backdrop, the study demonstrates how AI can help address the mounting administrative and cognitive burden in oncology. Rather than replacing clinicians, the technology is positioned as a decision-support tool that improves information processing and consistency.
More complete insights with AI
Researchers evaluated six open-source large language models developed by Meta, Google, Mistral AI and DeepSeek. The models analyzed 94 de-identified lung cancer pathology reports and generated structured summaries.
These AI-generated outputs were compared with physician-written summaries and assessed by a panel of oncologists on accuracy, completeness, conciseness and clinical risk. Across all models, AI consistently produced more comprehensive summaries, particularly in capturing molecular and genetic findings. These are data points that are critical for personalized treatment decisions.
DeepSeek and Llama 3.1
Among the evaluated systems, DeepSeek and Meta’s Llama 3.1 model delivered the strongest results. Notably, these models can be deployed locally, offering healthcare organizations greater control over data privacy and implementation.
The research team is currently developing a prototype application based on Llama 3.1, which would allow clinicians to upload pathology reports and receive AI-generated summaries for review. Further validation studies are required before clinical deployment.
The authors emphasize that AI should function as a complementary layer within clinical workflows. By structuring complex information and reducing the risk of overlooked details, AI can enable clinicians to focus more on patient care.
This is particularly relevant for patients with advanced or complex cancers, whose diagnostic histories may span dozens of pages. In such cases, missing a single actionable genetic marker can have significant implications for treatment.
Improving standardization with AI
The findings underscore the potential of AI to improve standardization and efficiency in clinical documentation. As oncology continues to evolve toward precision medicine, tools that can reliably synthesize complex datasets are likely to play an increasingly important role in supporting high-quality, data-driven care.
The Northwestern Medicine study, titled "Toward Automating the Summarization of Cancer Pathology Reports Using Large Language Models to Improve Clinical Usability," was published in JCO Clinical Cancer Informatics.
A couple of weeks ago, researchers from Flinders University showed that adding visual capabilities to AI-powered medical scribes can significantly improve clinical documentation accuracy. While current AI scribes rely mainly on audio, the study demonstrates that combining audio with visual input, using tools like Google Gemini and Ray-Ban Meta smart glasses, enables more complete capture of clinical details.
In simulated medication interviews, vision-enabled AI achieved 98% accuracy compared to 81% for audio-only systems, with major improvements in identifying medication type and dosage. The technology captures visual cues such as medicine packaging and body language, enhancing data quality.
