Fig 1.
Typical single-channel CNN architecture.
Fig 2.
Multi-channel Xception-based thyroid cancer detection framework.
Fig 3.
Single input dual-channel model.
Fig 4.
Double inputs dual-channel model.
Fig 5.
Double inputs dual-channel (four-channel) model.
Fig 6.
Ultrasound images from DDTI: TIRADS 2 on the left-side with well defined margins and no calcification, labeled as “benign”; TIRADS 5 on the right-side with micro-calcification, labeled as “malignant”.
Fig 7.
Segmented thyroid gland CT scans from patient No. 277: Left-side is Goiter (Benign) and right-side is Papillary Cancer (Malignant).
Table 1.
Data sets descriptions.
Table 2.
Comparison of various CNN models in binary classification tasks: Different sources of ultrasound images were applied for thyroid cancer detection.
Table 3.
Comparison of different CNNs on thyroid cancer diagnosis via CT.
Fig 8.
Initial binary classification task running time comparison for the 11 models.
Table 4.
Single-channel and dual-channel comparison.
Table 5.
DIDC and four-channel comparison.
Class 0 to 3 indicates the patient either has normal thyroid (0), has malignant left-side thyroid (1), has right-side thyroid malignant (2), or has both sides malignant (3).
Fig 9.
DIDC and four-channel 10-fold cross-validation accuracy results.
The average 10-fold scores for each epochs were demonstrated using the black line, and the average testing accuracy for DIDC is 0.95, for four-channel is 0.94.
Table 6.
Experiments comparison with existing literature.