Applications of Large Language Models in Breast Cancer Diagnosis : A Comprehensive Review of Techniques and Outcomes

Abstract

Breast cancer, with 2.3 million new cases annually, faces significant diagnostic challenges due to healthcare workforce shortages and prolonged assessment times. This systematic review evaluates large language models (LLMs) in breast cancer diagnosis, analyzing their potential to enhance diagnostic accuracy and healthcare efficiency. We conducted a comprehensive search of PubMed, SCOPUS, and Google Scholar databases for studies published through December 2024, focusing exclusively on LLM diagnostic applications. Our review examined diverse architectures including GPT-3.5, GPT-4, Claude, BERT variants, and PathologyBERT across pathological assessment, imaging interpretation, clinical information extraction, and BI-RADS classification. Fine-tuned models consistently outperformed general-purpose counterparts, achieving diagnostic accuracies exceeding 90%. GPT-4 demonstrated superior clinical scenario evaluation (quality score 3.56±0.55), while PathologyBERT achieved 73% accuracy in pathology text mining. However, performance declined with complex linguistic patterns and rare presentations, dropping up to 12% for ambiguous cases and to 78% for uncommon clinical scenarios. Despite promising results in routine applications, LLMs struggled with nested clinical information and nuanced language interpretation. The review highlights needs for standardized evaluation frameworks, diverse training datasets, and human-in-the-loop validation. While LLMs show significant promise for breast cancer diagnostic support, careful validation and responsible implementation are essential for safe clinical integration.

Ähnliche Arbeiten