Abstract
BACKGROUND: Accurate preoperative prediction of the invasiveness of lung nodules on computed tomography (CT) can avoid unnecessary invasive procedures and costs for low-risk patients. While previous studies approached this task using cross-sectional data, this study aimed to utilize the commonly available longitudinal data of lung nodules through sequential modelling based on long short-term memory (LSTM) networks. METHODS: We retrospectively included 171 patients with lung nodules that were followed-up at least once and pathologically diagnosed with adenocarcinoma for model development. Pathological diagnosis was the gold standard for deciding lung nodule invasiveness. For each nodule, a handful of semantic features, including size intensity and interval since first discovery, were obtained from an arbitrary number of CT scans available to individual patients and used as input variables to pre-operatively predict nodule invasiveness. The LSTM-based classifier was optimized by extensive experiments and compared to logistic regression (LR) as baseline with five-fold cross-validation. RESULTS: The best LSTM-based classifier, capable of receiving data from an arbitrary number of time points, achieved better preoperative prediction of lung nodule invasiveness [area under the curve (AUC), 0.982; accuracy, 0.924; sensitivity, 0.946; specificity, 0.881] than the best LR (AUC, 0.947; accuracy, 0.906; sensitivity, 0.938; specificity, 0.847) classifier. CONCLUSIONS: The longitudinal data of lung nodules, though unevenly spaced and varying in length, can be well modeled by the LSTM, allowing for the accurate prediction of nodule invasiveness. Given that the input variables of the sequential modelling consist of a few semantic features that are easily obtained and interpreted by clinicians, our approach is worthy further investigation for the optimal management of lung nodules.