AI Models Outperform Physicians in Summarizing Complex Cancer Pathology Reports

A study from Northwestern Medicine has found that artificial intelligence models can produce more comprehensive summaries of complex cancer pathology reports than physicians. The research, published on April 8, 2026, in JCO Clinical Cancer Informatics, suggests that AI may help alleviate the increasing cognitive burden on oncologists as diagnostic data becomes more detailed.

The study analyzed real-world lung cancer pathology reports from de-identified patients. Researchers tested six open-source AI models, including systems developed by Google, Meta, DeepSeek and Mistral AI, to evaluate their ability to synthesize complex medical information.

Addressing the Complexity of Modern Pathology

The need for automated summarization stems from the evolving nature of oncology. As biomarker testing expands and patient survival rates improve, pathology reports have become increasingly longitudinal and detailed.

These reports often span multiple medical institutions, requiring clinicians to synthesize vast amounts of data under significant time pressure to make critical treatment decisions.

According to Dr. Mohamed Abazeed, chair and professor of radiation oncology at Northwestern University Feinberg School of Medicine and senior author of the study, As cancer care becomes increasingly complex, the burden of synthesizing complex reports is growing rapidly.

Key Findings on AI Performance

The Northwestern Medicine study found that several open-source AI models consistently outperformed physician-written summaries in terms of completeness. The AI models were particularly effective at capturing molecular and genetic findings, which are essential for determining the appropriate course of treatment.

The research suggests that Large Language Models (LLMs) can efficiently convert multidimensional pathology results into coherent and concise summaries, potentially allowing physicians to dedicate more time to direct patient care.

Broader Context of AI in Diagnostics

While the Northwestern study highlights success in summarization, other research indicates a more varied landscape regarding the overall diagnostic performance of generative AI. A systematic review and meta-analysis of 83 studies published between June 2018 and June 2024 reported an overall diagnostic accuracy of 52.1% for generative AI models.

That analysis found no significant difference in performance between AI models and non-expert physicians. However, it did note that AI models performed significantly worse than expert physicians.

AI is being integrated into digital pathology workflows to support the interpretation of stains, such as haematoxylin and eosin, which assists in tumor classification and grading.

Remaining Challenges and Integration

Despite the potential for AI to improve the efficiency of report synthesis, several hurdles remain before widespread clinical deployment is fully realized. Experts have identified ongoing challenges in the following areas:

• The interpretability of AI-generated outputs.
• The rigorous validation of models in real-world settings.
• The seamless integration of these tools into existing clinical workflows.

The current evidence suggests that while generative AI demonstrates promising capabilities and may enhance medical education and healthcare delivery, it has not yet achieved the level of reliability associated with expert physicians.