Quantifying the Uncertainty in 3D CT Lung Cancer Images Classification

Lung cancer has caused the deaths of millions of people. Early diagnosis is proven vital to increase survival rate. This can be done using computed tomography screening, and when such a system is in place with a computer-aided diagnostic system and deep learning algorithms, it has been proven to be an effective approach. However, most existing systems do not quantify the uncertainty in the prediction. Radiologists should have complete confidence in the predictions during the screening. Therefore, to address this issue, we have developed an uncertainty-aware model framework not only to classify lung cancer nodules from 3D computed tomography images, but also to estimate the associated uncertainty in the prediction of the model. We have used the ResNet50 architecture as our base model and, through investigation with different loss functions and the inclusion of class weights, we concluded that the use of the focal loss function shows its superiority over the results of the benchmark model. Two approaches have been adopted to estimate the uncertainty during test time; Monte Carlo dropout and Test Time Augmentation, and from the multiple predictions obtained, three uncertainty metrics have been calculated; entropy, standard deviation, and Bhattacharyya coefficient. Our model achieves the best performance with the F1 score, 83.0% using the dropout rate, p, 0.5 and Monte Carlo iterations, M, 10. We have also shown that using a threshold value can greatly improve the classification performance even more, but the amount of data to be referred needs to be taken into account.