Breast cancer is the most frequent type of cancer among women. Since early diagnosis provides a better prognosis, different techniques have been developed by researchers all over the world. Several studies proved the efficiency of infrared image as a breast cancer screening technique. This paper proposes a methodology for analyzing infrared thermography of breast, considering distinct protocols, in order to classify patients images as healthy or non-healthy due to anomalies such as cancer. The major contribution of this work is to provide accurate classification using Convolutional Neural Networks, which were not exploited in previous works. Many methods relies on handcrafted features and traditional classificators, such as Support Vector Machines. We obtained competitive results compared to other works and we design an appropriate modelling which takes advantage of this type of deep learning architecture. Our proposal obtained 98% of accuracy for static protocol and 95% for dynamic protocol.
2 February 2023 in Computer Methods and Programs in Biomedicine
Breast cancer is the leading type of cancer among women. According to the World Cancer Research Fund, in 2018, over 2 million new cases were detected around the world. Despite its high occurrence, early detection provides a better prognosis and helps increases the patient’s survival. Significant advances in screening techniques, such as infrared imaging, have provided a cheap and less invasive way to detect the disease. Besides, computational tools can be used to assist doctors in providing a better diagnosis. Therefore, this paper presents a segmentation method based on U-Net Convolutional Neural Networks. In contrast to the state-of-the-art, machine learning approaches have been shown to be efficient for ROI segmentation, reaching an accuracy of 98.24% over the frontal view, and 93.6% over the lateral view. This segmentation method may be very useful for classification tasks, once the region of interest is well delimited for feature extraction.
Brain tumor corresponds to a group of diseases in which abnormal cells grow exponentially within the brain, the most important organ of the central nervous system. This disease is responsible for thousands of deaths worldwide. Early diagnosis and treatment increase the chance of survival. The most common technique to detect such tumors is analyzing MRI scans which carry subjectivity and lacks reproducibility. In order to improve the accuracy of such technique, a computer-aided system has been proposed, using Machine Learning and Image Processing to classify scans, characterizing such tumors, and acting as a second opinion. This paper benchmark several machine learning methods, such as ensemble and neural approaches, along with Radiomic features to classify MRI acquired on T1, T2, and FLAIR modalities, between healthy, glioma, meningioma, and pituitary tumor. Our experiments were conducted following K-Fold Cross-Validation and the best results were achieved by XGBoost and Deep Neural Network, in which XGBoost achieved an average accuracy of 88.51%and a sensitivity of 89.10 %, and the Deep Neural Network achieved 87.09% and a sensitivity of 87.38%. Such Computer Vision technique may assist radiologists to provide an accurate and efficient tumor diagnosis using MRI.