Computer-aided classification of lung nodules on CT scans via 3D CNNs.

Abstract

e18047 Background: Computer-aided diagnosis based on deep learning methodologies has demonstrated great potential to assist disease diagnosis with accuracy and efficiency. Specifically, the quantitative and qualitative analysis on lung nodules has proven to be important for the early-detection of lung cancer as well as its treatment in clinical practice. This study developed a 3D CNN model to facilitate the classification of pulmonary nodules. Methods: We collected 3956 lung CT scans (slice thickness≤3mm) with multiple lung nodules from 15 Class-A hospitals in China, 1155 lung CTs from Luna 16 dataset and Data Science Bowl 2017. There were 30 senior radiologists responsible for annotation and each CT scan was annotated by two of them randomly. Another 4 senior associate chief physicians were divided into two groups, each group was responsible for arbitration when conflicts occurred between the annotation doctors. All the annotated CTs were randomly selected and split to construct training, validation and test sets. We pre-processed the CTs and utilized 3D CNNs to classify these nodules as solid, partial-solid, ground glass opacity, calcified, pleural solid and pleural calcified. ROC analysis was used, and the classification capability was assessed by classification accuracy and the AUC score. Results: Table shows the overall results. The proposed model yielded an AUC score of 0.97 for the ground glass opacity and 0.90 for calcified nodules in the training set, while the AUC of them were 0.93 and 0.93 respectively in the validation set. For the test set, we got an AUC score of 0.94 for the ground glass opacity. The average classification time for each nodule was less than 0.005 sec. Conclusions: Our model may assist clinical diagnosis of lung cancer and increase its objectivity and accuracy, and the fast processing speed proves its feasibility to be applied in real clinical practice. In the future, we will enrich the dataset with clinical and genetic information, thus improving our model to boost its performance. [Table: see text]