Abstract:

Objective

The 3D Res U-Net-Faster RCNN model was meticulously trained and learned using CT imaging data of histopathology-confirmed pulmonary nodules to develop a predictive model capable of distinguishing between benign and malignant pulmonary nodules.

Methods

A retrospective study incorporating 528 pulmonary nodules cases （malignant 442 and benign 86 ） treated between October 2020 and October 2023 was conducted. The cases were randomly divided into a training set and a test set in a 7∶3 ratio. A modified 3D residual U-shaped network， fused with faster region-based convolutional neural networks （Faster R-CNN）， was utilized to identify regions of interest （ROI） within pulmonary nodules and to extract CT imaging features. This approach was employed to construct a predictive model capable of distinguishing between benign and malignant pulmonary nodules， as well as to judge and screen the CT imaging feature weights of malignant nodules. The diagnostic accuracy of the model for pulmonary nodules was determined by confusion matrix， accuracy， by precision， recall， F1 value， Dice similarity coefficient （dice loss）， subject operating characteristic curve（receiver operating characteristic， ROC）， and the performance of the model was verified with external data.

Results

In the pulmonary nodule properties prediction model based on 3D Res U-Net-Faster RCNN deep learning technology， the Dice Loss of segmentation ROI was 0.85， and the accuracy of the test set to identify malignant lung nodules was 0.85， the recall rate 0.76， F1 value 0.80， and the area under the curve （area under the curve， AUC） value 0.86. For the external validation set， lung nodules identification accuracy was 0.86， malignant nodules identification accuracy 0.92， recall rate 0.87， F1 value 0.90； benign nodules identification accuracy 0.92， recall rate 0.82， and F1 value 0.87. The mean gray value， maximum diameter to volume ratio， and surface area to volume ratio gave the highest weight to the prediction of malignant lung nodules. The diameter， burr and vascular travel were significantly different between the benign and malignant nodules （P＜0.05）.

Conclusion

The AI-driven diagnostic model constructed based on the CT-imaging features of deep learning of 3D Res U-Net-Faster RCNN technology has good predictive performance for the properties of pulmonary nodules and has auxiliary diagnostic significance for improving the screening of early lung cancer.

Key words: Pulmonary nodules, 3D residual u-net, Faster region-based convolutional neural networks, Predictive model