Banca de DEFESA: ANDERSON FELIX DA SILVA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : ANDERSON FELIX DA SILVA
DATE: 24/02/2022
TIME: 10:00
LOCAL: Remota
TITLE:

Intelligent system to support breast cancer diagnosis using thermographic imaging and deep artificial neural networks

KEY WORDS:

Diagnosis of breast cancer; mobile application; mammography; thermography; deep artificial neural networks.

PAGES: 95
BIG AREA: Engenharias
AREA: Engenharia Biomédica
SUMMARY:

Breast cancer is the deadliest form of cancer among women in both developed and underdeveloped and developing countries. Breast cancer mortality is directly linked to disease prevention strategies, such as educational campaigns and technologies to support early cancer diagnosis. The most used technique to support the diagnosis of breast cancer by imaging is X-ray mammography. However, mammography has its disadvantages, such as the cost, the use of ionizing rays (which may be related to cancer-causing factors), in addition to the discomfort in obtaining the image through breast compression. A complementary technique to mammography is breast thermography, which is based on the metabolic changes resulting from the appearance of altered cells in the breast tissue, which result in changes in heat distribution. Thermography has been proposed as a complementary technique to mammography, being more efficient compared to breast examination and serving as a screening system, allowing early detection of breast lesions and reducing mortality. That said, this work aims to analyze the use of deep neural networks together with different classification techniques, for the recognition of lesions in thermographic images using the Weka machine learning software. As well as substantiate a model that can be explored in applications to support the diagnosis of breast cancer for classification of lesions in thermography. Initially, in the attribute extraction step, different deep networks from Weka's DeepLearning4j library were used: LeNet, ResNet50, NASNetMobile, SqueezeNet and Inception v3. Soon after, the selection of the best attributes was performed using PSO. Then, in the classification and training stage, 30 experiments were generated for the following classifiers: Naive Bayes, Bayes Net, Random Tree, J48 Decision Tree, Random Forest, SVM Support Vector Machine and MLP Multilayer Perceptron Network. The results were compared using boxplots and tables for the metrics of Accuracy, Kappa Index, Sensitivity, Specificity, Area Under the ROC Curve and Training Time (ms). Finally, the best performances between the deep neural networks and the classifiers used were analyzed. In addition, the performance before and after the selection of attributes was also analyzed, in order to determine the most efficient model to be used. As a result, the Inception V3 deep network combined with the SVM classifier with a polynomial kernel of 3 had the highest accuracy rate for the approach without feature selection, obtaining 79.92%. However, using the selection of attributes, the CNN Inception V3 combined with the SVM classifier with a polynomial kernel of 4, obtained an accuracy of 78.55% with a training time twice as short.

BANKING MEMBERS:
Interno - 1170977 - RICARDO EMMANUEL DE SOUZA
Interno - 1807632 - WELLINGTON PINHEIRO DOS SANTOS
Externa ao Programa - 1132340 - RITA DE CASSIA FERNANDES DE LIMA