A Bayesian neural network model based on CT images for staging non-small cell lung cancer

Automatic staging of non-small cell lung cancer based on CT images is an important auxiliary diagnostic method for lung cancer. Many deep learning methods have been proposed for this task, but they cannot measure the confidence of model predictions. Therefore, it is necessary to develop new deep learning methods that can measure the uncertainty of the model predictions. In this paper, we propose a Bayesian neural network (BNN) model based on CT images for staging non-small cell lung cancer patients, which can measure and utilize the uncertainty of the model predictions. The entire workflow consists of two stages: image-level classification and patient-level classification. Unlike conventional neural networks, the model's trainable pa-rameters (or weights) are set to be random, so for a fixed input, the model outputs a distribution of predictions rather than a unique result. Based on the uncertainty measurement results, two post-processing methods for model refinement are developed: 1) investigating the correlation between high uncertainty and misclassification and selecting a low uncertainty subset on the validation set to prevent the model from being overconfident, and 2) weighting and integrating the image prediction scores to obtain patient staging results. The results of training and validation on the public dataset Lung-PET-CT-Dx show that the proposed BNN model and its refined versions outperform the Cascade NN model and the RepVgg model based on transfer learning in terms of prediction accuracy and area under the receiver operating characteristic curve (AUROC) metrics. The BNN model achieved an average accuracy and AUROC of 0.760 and 0.836, respectively, with improvements 0.030 and 0.055 over Cascade NN and 0.080 and 0.067 over RepVGG, respectively. In addition, the two model refinement methods improved the performance to 0.792 and 0.842, and 0.770 and 0.841, respectively.