Fig 1.
Breast thermography procedure (thermal image is aquired at room temperature = 22°C).
Fig 2.
Flowchart of the proposed method.
Fig 3.
Example of U-Net architecture [33].
Fig 4.
Example of breast area segmentation with U-Net.
Fig 5.
Architecture of the proposed deep learning model.
Fig 6.
Different cases of breast (a) small breast (b) large breast (c) asymmetric breast.
Table 1.
Dataset description.
Fig 7.
Breast area segmentation resuls (a) thermal image (b) ground truth (c) output.
Fig 8.
The training progress of the proposed deep learning model.
Fig 9.
The confusion matrix of the proposed model.
Table 2.
Comparison between solvers (initial learn rate = 2.0e−3, number of epochs = 30 and batch size = 60).
Table 3.
The impact of using different number of epochs on the classification accuracy, sensitivity and specificity (solver = ADAM, initial learn rate = 2.0e−3, batch size = 60).
Table 4.
Impact of using different batch size on the classification accuracy, sensitivity and specificity (solver = ADAM, initial learn rate = 2.0e−3, number of epochs = 30).
Table 5.
Impact of starting the training process with different initial learn rate on the classification accuracy, sensitivity and specificity (solver = ADAM, batch size = 60 and number of epochs = 30).
Table 6.
Comparison between the performance metrics of different CNN models and the proposed model.
Fig 10.
Evaluation metrics over different dataset size.
Table 7.
Comparison between the performance metrics of different machine learning classifier with texture features and the proposed model.
Table 8.
Comparison between the performance metrics of different machine learning classifier with HOG features and the proposed model.
Table 9.
Results of the ANOVA test of the proposed model and CNN models.
Table 10.
Comparison with other studies on breast cancer detection (n = normal, ab = abnormal, Ea = Early, Ac = Acute).
Table 11.
Table of abbreviation.