Academia.eduAcademia.edu

Figure 1 – uploaded by Maíra Santana

See full PDFdownloadDownload figure
Fig. 1 Proposed method for the identification and classification of breast lesions in breast thermography images. Initially the images are converted to gray scale, then we extract features through DWNN method. Finally, the feature vectors are assessed by different classification algorithms to identify the existence of a breast lesion in the image and indicate the probable type of lesion.

Figure 1 Proposed method for the identification and classification of breast lesions in breast thermography images. Initially the images are converted to gray scale, then we extract features through DWNN method. Finally, the feature vectors are assessed by different classification algorithms to identify the existence of a breast lesion in the image and indicate the probable type of lesion.

Related Figures (8)

‘ig. 2 Examples of images from each of the possible classes in the breast thermography database. In the ipper left corner is an example of the uninjured class; right beside it, in the upper right corner, is an image ‘ontaining a cystic-type lesion. The lower left and right images are, respectively, of benign and malignant acinane
‘ig. 2 Examples of images from each of the possible classes in the breast thermography database. In the ipper left corner is an example of the uninjured class; right beside it, in the upper right corner, is an image ‘ontaining a cystic-type lesion. The lower left and right images are, respectively, of benign and malignant acinane
Fig. 3 Positions used in performing the breast thermography exam. In the upper left portion are T1 and T2, which are frontal images of both breasts, with the arms in different positions. MD shows the frontal image of the right breast only, while ME shows the left breast in the same position. The lower images are from lateral views, with LEMD and LEME being images of the external side of the right and left breasts, respectively; LIMD and LIME correspond to the internal lateral view of each breast.
Fig. 3 Positions used in performing the breast thermography exam. In the upper left portion are T1 and T2, which are frontal images of both breasts, with the arms in different positions. MD shows the frontal image of the right breast only, while ME shows the left breast in the same position. The lower images are from lateral views, with LEMD and LEME being images of the external side of the right and left breasts, respectively; LIMD and LIME correspond to the internal lateral view of each breast.
Fig. 4 Preparation procedures for the acquisition of thermographic images
Fig. 4 Preparation procedures for the acquisition of thermographic images
Fig. 5 Mechanical apparatus used in the acquisition room. In (1) are the rails used to move the camera support (2) towards the patient, positioned on the swivel chair shown in (3); the bars in (4) are used for positioning the arms during the exam. In this work, we used only the Tl and T2 images, as this positioning favors the visualization of both breasts and the identification of possible anomalies and issues ‘elated to asymmetries. Thus, 336 images were used, distributed in 4 classes accord- ing to what is presented in Table 1. Since the amount of images per class is different, 1 class balancing step was added to avoid classification biases, since, with unbalanced oases, algorithms tend to favor classes with more representatives over less populated classes. In this sense, the balance is made from the insertion of synthetic instances,
Fig. 5 Mechanical apparatus used in the acquisition room. In (1) are the rails used to move the camera support (2) towards the patient, positioned on the swivel chair shown in (3); the bars in (4) are used for positioning the arms during the exam. In this work, we used only the Tl and T2 images, as this positioning favors the visualization of both breasts and the identification of possible anomalies and issues ‘elated to asymmetries. Thus, 336 images were used, distributed in 4 classes accord- ing to what is presented in Table 1. Since the amount of images per class is different, 1 class balancing step was added to avoid classification biases, since, with unbalanced oases, algorithms tend to favor classes with more representatives over less populated classes. In this sense, the balance is made from the insertion of synthetic instances,
Table 2 Classifiers parameters: SVM with linear kernel; MLP with 100 neurons in the hidden layer; ELM with sigmoid kernel; random forests with 100 trees; standard random tree and J48; and standard Bayesian network and Naive Bayes classifiers. After extracting attributes, the set was submitted to a classification step. We evalu- ate the classification performance of different methods, the Bayesian network (Bayes Net), the naive Bayes classifier (Naive Bayes), the Multilayer Perceptron (MLP), the Support Vector Machine (SVM), the Extreme Learning Machines (ELM), in addi- tion to the search tree-based classifiers: J48, Random Tree and Random Forest. The settings for each classifier are presented in Table 2.
Table 2 Classifiers parameters: SVM with linear kernel; MLP with 100 neurons in the hidden layer; ELM with sigmoid kernel; random forests with 100 trees; standard random tree and J48; and standard Bayesian network and Naive Bayes classifiers. After extracting attributes, the set was submitted to a classification step. We evalu- ate the classification performance of different methods, the Bayesian network (Bayes Net), the naive Bayes classifier (Naive Bayes), the Multilayer Perceptron (MLP), the Support Vector Machine (SVM), the Extreme Learning Machines (ELM), in addi- tion to the search tree-based classifiers: J48, Random Tree and Random Forest. The settings for each classifier are presented in Table 2.
Fig. 6 Results of (a) accuracy and (b) kappa index for the thermography image base using attributes extracted by the Deep-Wavelet Neural Network
Fig. 6 Results of (a) accuracy and (b) kappa index for the thermography image base using attributes extracted by the Deep-Wavelet Neural Network
Table 4 Confusion matrix for best and worst results using Breast Thermography database
Table 4 Confusion matrix for best and worst results using Breast Thermography database
Table 5 Sensitivity, specificity and efficiency metrics of the approach using the breast thermography basis 5 Conclusion EEE From the results, it is observed that the deep network approach wit! h DWNN was able to represent the images well. When associated with support vector machines (SVM), the identification of lesions in breast thermograms reached around 99% and more than 0.95 kappa index. Good results were also an accuracy obtained with MLP, Random Forest and ELM, indicating that the problem can be generalized, but often in a non-linear way. The good performance of these methods ex results obtained with decision trees (J48 and Random Tree), since t commonly achieve good results when the basis is very specific. The performances with Bayesian methods point to a considerable depend the attributes. plains the low hese methods relatively low ency between
Table 5 Sensitivity, specificity and efficiency metrics of the approach using the breast thermography basis 5 Conclusion EEE From the results, it is observed that the deep network approach wit! h DWNN was able to represent the images well. When associated with support vector machines (SVM), the identification of lesions in breast thermograms reached around 99% and more than 0.95 kappa index. Good results were also an accuracy obtained with MLP, Random Forest and ELM, indicating that the problem can be generalized, but often in a non-linear way. The good performance of these methods ex results obtained with decision trees (J48 and Random Tree), since t commonly achieve good results when the basis is very specific. The performances with Bayesian methods point to a considerable depend the attributes. plains the low hese methods relatively low ency between
Related topics:
Computer ScienceArtificial Intelligencebreast ultrasound
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved