Fig 1.
The four stages of lung cancer development.
Fig 2.
The constituent steps of the proposed method.
Fig 3.
The 3D-GLCMs are computed from VSs of nodule volume and fed to a recurrent neural network.
Fig 4.
In 2.5D-GLCM mode the GLCMs computed for each slice are fed to the recurrent neural network.
Table 1.
The structure of the proposed LSTM-based fusion method.
f refers to the number of elements in the extracted features.
Table 2.
The comparison of WSA-Otsu with counterpart methods.
Fig 5.
The convergence curves of Otsu method in terms of fitness values at five different levels.
Fig 6.
The thresholding time of the proposed method vs. other counterparts.
Fig 7.
ROC curves for three fusion models and their corresponding macro-averaged AUC values.
Fig 8.
ROC curves for three fusion models and their corresponding micro-averaged AUC values.
Table 3.
A comparison between the proposed 2D and 2.5D GLCM fusion methods and other deep CNN-based methods regarding the accuracy, sensitivity, and specificity metrics.
Table 4.
A comparison between the proposed fusion methods for different modes of GLCM computation and the recently proposed classification methods regarding the accuracy, sensitivity, and specificity metrics.
Fig 9.
ROC curves for baseline and proposed models, and their corresponding micro-averaged AUC values.
Fig 10.
ROC curves for baseline and proposed models, and their corresponding macro-averaged AUC values.
Table 5.
The classification results in terms of micro-averaged AUCs trained on the LIDC-IDRI dataset and tested on LIDC-IDRI and LUNGx datasets for the proposed and baseline models.
Table 6.
The T-test p-values obtained comparing the average AUC of the LSTM-based deep fusion of 3D-GLCMs with other baseline models.
Table 7.
Results of the ablation study for the type of features presented to the LSTM-based deep fusion model.
Fig 11.
The training time of different models for five epochs.
Fig 12.
The test time of different models.