Identification of benign and malignant pulmonary nodules on chest CT using improved 3D U-Net deep learning framework.

PURPOSE

To accurately distinguish benign from malignant pulmonary nodules with CT based on partial structures of 3D U-Net integrated with Capsule Networks (CapNets) and provide a reference for the early diagnosis of lung cancer.

METHOD

The dataset consisted of 1177 samples (benign/malignant: 414/763) from 997 patients provided by collaborating hospital. All nodules were biopsy or surgery proven, and pathologic results were regarded as the "golden standard". This study utilized partial U-Net to capture the low-level (edge, corner, etc.) information and CapNets to preserve high-level (semantic information) information of nodules. For CapNets, each capsule had a 4 × 4 matrix representing the pose and an activation probability representing the presence of an object. Furthermore, we chose accuracy (ACC), area under the curve (AUC), sensitivity (SE) and specificity (SP) to evaluate the generalization of the proposed architecture and compared its identification performance with 3D U-Net and experienced radiologists.

RESULTS

The AUC of our architecture (0.84) was superior to that (0.81) of the original 3D U-Net (p = 0.04, DeLong's test). Moreover, ACC (84.5 %) and SE (92.9 %) of our model were clearly higher than radiologists' ACC (81.0 %) and SE (84.3 %) at the optimal operating point. However, SP (70 %) of our model was slightly lower than radiologists' SP (75 %), which might be the result of class imbalance with limited benign samples involved for algorithm training.

CONCLUSIONS

Our architecture showed a high performance for identifying benign and malignant pulmonary nodules, indicating the improved model has a promising application in clinic.