Predicting future lung cancer risk in low-dose CT screening patients with AI tools

Radiologists are increasingly recognizing the challenge of “interval lung cancers” (ILCs), which emerge after a negative screening CT scan but before the next scheduled scan in 1-2 years. Unlike cancers detected during the baseline scan, ILCs are often diagnosed at a late stage, more aggressive, hence less treatable, and tend to have poorer outcomes. This study aims to develop an artificial intelligence (AI) tool that can predict the future risk of cancer by identifying subtle features in low-dose CT (LDCT) images of “normal-appearing” lungs before the onset of interval cancer. A convolutional neural network (CNN) was used to extract key features from both normal and abnormal regions in the scans of 36 ILC patients from two retrospective studies (Training set: 26 patients, Test set: 10 patients). We collected CT scans from the first scan where ILC was identified (denoted t0CT) and from prior scans (denoted pCT) taken during different screening intervals. To establish a volume of interest (VOI), we placed bounding boxes around the ILC nodule in the t0CT scan. A same-sized VOI was then asymmetrically placed in the contralateral lung, which showed no signs of ILC, serving as a control. A nonparametric statistical test was used to rank the discriminative power of the 128 features from the final layer of the trained CNN. A support vector classification (SVC) model, trained using the 7 highest-ranked individual features, achieved an AUC of 0.87, sensitivity of 0.82, specificity of 0.82, and an F1-score of 0.75. This AI-based approach has the potential to predict the risk of future ILCs during earlier screening rounds, even when the lungs appear visually normal in the routine LDCT scans.