Pulmonary embolism (PE) presents a significant diagnostic challenge, necessitating clear and precise communication between radiologists and referring physicians to optimize patient outcomes. Variability in the language used to convey diagnostic uncertainty in computed tomographic pulmonary angiography (CTPA) reports can exacerbate ambiguity, potentially leading to inappropriate clinical management and compromised patient care. Addressing this issue, a standardized way to measure diagnostic uncertainty in CTPA reports holds promise for enhancing communication precision, reducing treatment errors, and ultimately advancing the quality of care for patients with suspected PE.
This webinar explores advanced natural language processing (NLP) techniques for assessing PE diagnostic uncertainty in CTPA reports, particularly in low-resource settings with limited annotated data. Drawing from a comprehensive analysis of annotated radiology reports, this talk will discuss the performance of multiple methodologies, including prompt-dependent large language models (LLMs) like GPT-3.5 fine-tuning and in-context learning, as well as a prompt-free framework, Sentence Transformers Fine-tuning (SetFit), specially designed for few-shot learning. In addition to demonstrating the potential of these models to unlock contextual semantics in radiology reporting, a novel interpretability analysis is introduced, providing insights into radiologists' preferences for model transparency and reasoning explainability.
Presenter
