Lung-RADS + AI: A Tool for Quantifying the Risk of Lung Cancer in Computed Tomography Reports

Cancer represents a significant public health challenge. In 2023, it is projected that cancer will cause the loss of thousands of lives in the United States alone. Lung cancer is the second most commonly diagnosed cancer globally. It represents the deadliest form of malignant neoplasm, resulting in around 1.8 million fatalities in 2020. Chest computed tomography has found extensive application in the assessment of lung cancer, serving as a pivotal tool for non-invasive in vivo initial diagnosis. The utilization of guidelines within screening programs carries substantial significance as it seeks to reduce the requirement for excessive follow-up examinations. The Lung-RADS is one of these guidelines used for screening and follow-up of suspected Lung lesions. It introduces a classification of lung lesions as numbered categories according to their characteristics and degree of suspicion of malignancy. In this context, the main objective of this work is to assess the effectiveness of three Named Entity Recognition (NER) techniques, including CNN, BiLSTM and BERT variants to extract characteristics of pulmonary nodules in chest CT reports in Portuguese idiom and calculate the probability of malignancy index using the Lung-RADS guideline. Our top-performing model was BiLSTM-CRF, and this model achieved an average precision of 96%, an average recall of 88%, and an average f1-score of 90%. This research contributes to advancing NLP applications in the healthcare domain, specifically for Portuguese language radiology reports and demonstrates the potential for automated information extraction to enhance the accuracy and efficiency of lung cancer diagnosis and management.