A machine learning-based PET/CT model for automatic diagnosis of early-stage lung cancer

ObjectiveThe aim of this study was to develop a machine learning-based automatic analysis method for the diagnosis of early-stage lung cancer based on positron emission tomography/computed tomography (PET/CT) data.MethodsA retrospective cohort study was conducted using PET/CT data from 187 cases of non-small cell lung cancer (NSCLC) and 190 benign pulmonary nodules. Twelve PET and CT features were used to train a diagnosis model. The performance of the machine learning-based PET/CT model was tested and validated in two separate cohorts comprising 462 and 229 cases, respectively.ResultsThe standardized uptake value (SUV) was identified as an important biochemical factor for the early stage of lung cancer in this model. The PET/CT diagnosis model had a sensitivity and area under the curve (AUC) of 86.5% and 0.89, respectively. The testing group comprising 462 cases showed a sensitivity and AUC of 85.7% and 0.87, respectively, while the validation group comprising 229 cases showed a sensitivity and AUC of 88.4% and 0.91, respectively. Additionally, the proposed model improved the clinical discrimination ability for solid pulmonary nodules (SPNs) in the early stage significantly.ConclusionThe feature data collected from PET/CT scans can be analyzed automatically using machine learning techniques. The results of this study demonstrated that the proposed model can significantly improve the accuracy and positive predictive value (PPV) of SPNs at the early stage. Furthermore, this algorithm can be optimized into a robotic and less biased PET/CT automatic diagnosis system.

Automated summary

This study aimed to develop a machine learning-based automatic analysis method for the diagnosis of early-stage lung cancer using PET/CT data. A retrospective cohort study was conducted using PET/CT data from 187 cases of lung cancer and 190 cases of benign pulmonary nodules. The performance of the machine learning-based PET/CT model was tested and validated in two separate cohorts.