Fen Bilimleri Enstitüsü

Bu topluluk için Kalıcı Uri

http://hdl.handle.net/11527/8
İTÜ Bünyesinde lisansüstü yönetmeliklere uygun olarak çağdaş bilim ve teknolojinin gelişmesini izleyip bunları ülkemizde uygulama alanına aktarabilecek araştırma niteliği kazanmış yüksek lisans ve doktora öğrencisi yetiştirmek üzere gerekli faaliyetlerini sürdürmektedir.

Gözat

Yazar "Abdolsaheb Yılmaz, Tuba" ile Fen Bilimleri Enstitüsü'a göz atma
  • Öge
    Microwave spectroscopy based breast cancer diagnosis using support vector machines
    (Institute of Science and Technology, 2020-07-16) Önemli, Emre ; Akduman, İbrahim ; Abdolsaheb Yılmaz, Tuba ; 504171403 ; Biomedical Engineering ; Biyomedikal Mühendisliği
    Interactions of electromagnetic (EM) fields with materials relies on their intrinsic dielectric properties. Knowledge of the dielectric values of each material allows to develop electromagnetic technologies in many fields including medical technologies. There are a variety of electromagnetic medical technologies such as Microwave Imaging, Electrical Impedance Tomography and radiofrequency ablation and they promise faster, safer and low-cost applications. They rely on inherent differences among the dielectric properties of various biological tissue groups and health conditions. Hence, knowledge of the tissue dielectric properties of different biological tissues is crucial for developing EM healthcare technologies. Many works have been performed to investigate difference between dielectric properties of healthy and malignant tissues. It has been discovered that healthy and malignant tissues differ for the EM interactions because of the disperancies in their dielectrical properties. This contrast have been attributed to more water presence in malignant tumors. Breast carcinoma became one of the most researched cancer because of its high incidence and mortality rate. It is responsible for twenty three percent of new cancers and fourteen percent of cancer deaths in total. Thus, early diagnosis of the breast cancer is gaining more importance. Currently, there are some diagnostic methods such as mamography or MRI. However, they have some drawbacks such as harmful effects and low accuracy. Lately, microwave imaging (MWI) gained many interests. MWI fundamentally relies on the inherent dielectric contrast between healthy and malignant tissues. In cancer resection surgeries, determination of clear surgical margins is also possible using dielectric properties. Numerous studies were performed to expand the knowledge of the dielectric properties. However, existing dielectric datasets do not include every tissue type, frequency and temperature. Hence, more studies are needed. Open-ended coaxial probe has became the most preferred measurement method, because it is non-destructive, easy and suitable for biological materials. More dielectric data requires fast and accurate classification methods. For medical applications, most preferred one is Support Vector Machines (SVM). Being a supervised classification method, SVM is widely used because of its high classification performance on medical data. In this study, performance of SVM and infinite feature selection was investigated on the dielectric data of female rat normal breast tissues and malignant tumors in microwave frequencies. Measurements were conducted between 0.5 GHz and 6 GHz with 0.55 GHz intervals at 101 frequency points. Relative permittivity, conductivity and combination of them were tested separately. Firstly, they were tested without feature selection, raw dielectric data was also compared with normalization and logarithm of the dielectric data. Raw permittivity and combined data outperformed others resulting in 100% accuracy. Note that cross validation (CV) technique does not allow memorization of the learning model. Selecting top 100 features, the algorithm resulted in 100% accuracy with permittivity data whereas using top 50 features, it resulted in 99.23% accuracy with combined data. Using nested cross validation, features were selected as top 1 to top 100. Raw permittivity data gave more than 99% accuracy for more than sixty features. Using only one feature, 83.69% accuracy was obtained. Logarithm of the conductivity data resulted in 90.31% and 90% accuracy using one feature with linear and RBF kernels respectively. Best result of conductivity data is 98% using raw data and selecting top 70 features. With one feature, frequency of 5.505 GHz resulted in the best result. S11 response was also tested to avoid dielectric property calculation and to design narrow band devices. Note that this response indicates the energy transfer between probe and biological tissue related to tissue intrinsic electrical properties. Logarithm of the data outperformed with 93.85% accuracy using 10-fold linear SVM. Feature selection step was performed with 10-fold CV. With top 100 features, logarithm of data resulted in slightly higher performance as 91.85% accuracy with RBF kernel. With top 50 features, raw data was slightly better with 85.85% accuracy using linear SVM. Nested CV was applied to logarithm of S-parameter data. Selecting top 10 to 100, with decreasing number of features, accuracy dropped from 91.69% to 87.23% for RBF kernel and 91.38% to 87.08% for linear kernel. Besides, using top 1 to 10 features, accuracy dropped from 87.23% to 86.92% for RBF kernel and 87.08% to 83.08% for linear kernel. Best feature was corresponding to real part of S11 response at 610 MHz. The results show that dielectric measurement data can become acceptable diagnostic tool for breast cancer diagnosis. Thus, development of the EM medical technologies requires more tissue dielectric data. This study provides more dielectric data to the literature and it provides a perspective for analysing the dielectric data on the classification manner.