Vibration-based system design for early detection of breast tumors

According to the World Health Organisation (WHO), breast tumours are one of the most common diseases among women worldwide. Thanks to the early diagnosis of this disease, a 99% survival rate is achieved, which constitutes approximately 64% of the related cases. Therefore, in this study, a microcontroller-based vibration meter device has been designed for early diagnosis of breast tumour. With this device, constant frequency vibration excitation was applied on healthy and tumorous phantom breast models in the laboratory environment and vibration signals at certain distances were measured with the ADXL345 accelerometer sensor in MEMs. Before the analysis of the data set formed by the vibration signals obtained, the resonance frequency, which has the same effect at all points of the relevant breast phantom models and which is the frequency value at which the best analysis can be performed, was determined around 164 Hz. using the COMSOL program using the finite element model (FEM). Thus, 160 Hz. of the fixed frequency vibration excitations in the range of 100-200 Hz. applied to the breast phantom models was focused on and the vibration signals obtained around the resonance frequency were analysed. In the analysis phase of the vibration signals in the data set, firstly in the pre processing phase, a 5 Hz. high-pass IIR filter was used in order not to lose the low frequencies in the vibration signals and at the same time to filter the DC components. After pre-processing, in the stage of analysing the vibration signals and extracting critical features, time domain analyses of the related signals were performed by statistical methods such as RMS, standard deviation, skewness, kurtosis, median frequency methods. When the values obtained by these methods were analysed, it was observed that the RMS values of the unfiltered signals were prominent in distinguishing between healthy and tumour tissues. In addition, the median analysis values of the filtered signals showed that the distinction between healthy and tumorous can be made clearly. Following the time domain analyses, frequency domain analyses were performed with Fast Fourier Transform (FFT), Welch, Wavelet Packet Transform(WPT), Hilbert Huang Transform (HHT) and Empirical Mode Decomposition (EMD) methods. In the analyses performed with the FFT method, it was observed that the peak frequency values of the vibration signals obtained from healthy models were lower than the peak frequency values of the vibration signals obtained from tumour models. These values were also supported by maximum frequency analyses. In the analyses performed with the Welch method, it was observed that the peak frequency values of the vibration signals obtained from healthy models were lower than the peak frequency values of the vibration signals obtained from tumour models. Wavelet Packet Transform (WPT),another method used in the frequency domain analyses of the study, was used to examine the wavelet packets of the vibration signals at levels 3, 4 and 5. Accordingly, when the signal values obtained from the wavelet packets of the vibration signals obtained from healthy phantom models at level 3 were examined, it was observed that the filtered vibration signals were concentrated in the frequency range of 62,5-125 Hz.at level 3 packet 1 (𝑊3,1), while the vibration signals obtained from tumoured phantom models were concentrated in the frequency range of 125-187,5 Hz. at level 3 packet 2 (𝑊3,2). When the signal values obtained in the wavelet packets of the WPT at level 4 are examined, it is seen that the filtered vibration signals are concentrated in the 4th level 4th packet (𝑊4,4), 125-156,25 Hz. frequency range, while the vibration signals obtained from the tumoured phantom models are concentrated in the 4th level 7th packet (𝑊4,7) and 4th level 6th packet (𝑊4,6), 218,75-250 Hz. and 187,5-218,75 Hz. frequency ranges respectively. In Level 5 WPT analyses, it was observed that the filtered vibration signals from healthy phantom models were concentrated in the frequency range of 171,875-187,5 Hz. at Level 5 Pack 11 (𝑊5,11), while the vibration signals from tumoured phantom models were concentrated in the frequency range of 234,75-250 Hz. at Level 5 Pack 15 (𝑊5,15). The results show that WPT has a clear discriminating feature about the frequency characteristics of healthy and tumorous tissues. Hilbert Huang Transform (HHT) was also used in the frequency domain analyses of the doctoral thesis study. In the frequency domain analysis of the vibration signals obtained from healthy and tumorous breast phantom models with HHT, a difference was observed between the region where the frequency values are concentrated in the frequency-amplitude graphs of healthy tissues and the values where the frequency values are concentrated in the frequency-amplitude graphs of tumorous tissues. In addition, the Internal Mode Functions (IMF) were decomposed at 10 levels, and it was observed that the EMD analysis of the vibration signals at 10 levels was not effective in distinguishing between healthy and tumorous tissues. Feature vectors were created with the values obtained from the time and frequency domain analyses of the study and these feature values were used as inputs to the classification algorithms. Chi-Square and F-Score methods were used to determine the most discriminative features for the vibration signals obtained from healthy and tumour phantom models. By decomposing the WPT at levels 3, 4 and 5 and the EMD at 10 levels, 3 different feature vectors were analysed with these methods and the most discriminative features were determined. Accordingly, in both tests, the most discriminative features were obtained in the frequency range of 62,5-125 Hz. with 𝑊3,1 packet, in the frequency range of 125-156,25 Hz. with 𝑊4,4 packet, and in the frequency range of 171,875-187,5 Hz. with 𝑊5,11 packet. Finally, the most discriminative features were used as the inputs of Adaboost, MLPNN, SVM, KNN and LR classification algorithms and the most successful methods to perform healthy-tumour phantom model discrimination were evaluated. Accordingly, it was determined that the highest accuracy rate in the discrimination of healthy and tumorous tissues was 86.33% with the features obtained at Level 3 by wavelet packet transform of the filtered signals and these features were the values belonging to the feature numbers between 24 and 32 in the feature vector obtained by Level 3 WPT and obtained by the Adaboost classification algorithm. In addition, it was determined that the classification algorithms with the highest success rates were Adaboost with 82.66% accuracy rate, MLPNN with 82.33% accuracy rate and Logistic Regression with 80.66% accuracy rate.