Towards automating lung-rads classification in clinical routine insights from portuguese radiology reports

Abstract

Lung cancer has the highest mortality rate among all cancer types, affecting both men and women. It is estimated that lung cancer accounts for 21% of cancer deaths in each gender worldwide. This alarming statistic highlights the significant impact of lung can cer on overall cancer mortality, underscoring the urgent need for effective prevention, early detection, and treatment strategies to combat this disease. Lung cancer screening is a process designed to detect lung cancer in at-risk individuals, particularly those with a history of smoking. It involves annual low-dose computed tomography (CT) scans, careful interpretation of results, and timely follow-up to ensure early detection and treat ment. Several professional societies, including the American College of Radiology and the Fleischner Society, have published guidelines for the management of patients with pulmonary nodules detected during lung cancer screening. The guidelines are an important tool in screening programs aimed at reducing the incidence of unnecessary follow-up examinations and guiding optimal patient management. The Lung Computed Tomography Screening Reporting & Data System (Lung-RADS) is a standardized classification system for pulmonary nodules detected on imaging examinations, such as CT scans. Lung-RADS assesses the risk of malignancy (cancer) in these nodules and guides subsequent management decisions. In this context, this work aims to analyze the effectiveness of deep learning and large language models in extracting features of pulmonary nodules from Portuguese CT reports to enable automated classification of Lung-RADS. This work evaluated the effectiveness of BiLSTM-CRF, BioBERTpt, Gemini 1.5 Flash, GPT-4o, and Llama 3 70B. The results suggest that the Gemini 1.5 Flash stands out as the most effective model, outperforming the others in four of the five Lung-RADS classifications in the test set, with a weighted F1-score of 0.95, highlighting its effectiveness in accurately assessing lung nodules in various classification scenario.