Reconhecimento de entidades características para classificação de nódulos pulmonares em laudos médicos

Abstract

Lung cancer is the deadliest type of malignant neoplastic disease, causing 1.8 million victims in 2020. In absolute numbers, it is higher than the second and third leading types combined (colo rectal and liver cancer). Studies show that early diagnosis is crucial to increase the chances of successful treatment, primarily achieved through the detection of lung nodules via lung compu terized tomography (CT) scans. As a typical practice, the results of these scans are data stored in an unstructured, free text format, after passing through a storing process that contributes to its corruption, such as grammatical errors, typos, or even lack of convention of clinical terms. This lack of data structure creates an unnecessary obstacle for the disease diagnosis, as it is necessary to search in these results for the presence of nodules with malignant characteristics. Aiming for identifying and extracting relevant data for the disease diagnosis, a deep-learning model was trained to perform a named entity recognition task, enabling the retrieval of entities related to the nodules aspects from textual lung CT scan reports. The entities were defined taking in con sideration the characteristics needed to enable the classification according to the international Lung-Rads guideline. Six entities were defined to identify the quantity, type, size, location, and general characteristics of the nodules, as well as the presence of pulmonary emphysema. The spaCy framework was used to train the model, employing a BERT pre-trained language model architecture in Brazilian Portuguese. The model achieved a F1 score of 91.89% using a dataset of 600 documents provided by Unimed Maceio Hospital, manually annotated as described in the methodology.