The Research On Microwave-Based Breast Cancer Detection Algorithm

The incidence of cancer has grown rapidly in recent years.The ratio of breast cancer occupied among the cancers also increases year by year.The lat-est statistic shows that breast cancer has rank first in the incidence of the cancer among women,replaced uterine cancer.The increasing ratio of incidence of breast cancer is double times the average ratio of the world.However,if the breast cancer can be detected and cured in the early stage,the five-year sur-vival rate of one patient would be larger than 90%,according to the statistics from Canada Cancer Society.Therefore,the study of breast cancer detection is worthy to research from the view of both theoretical and practical.Magnetic Resonance Imaging(MRI),X-ray Mammography and Ultra-sound imaging are three most widely used breast cancer detection methods in clinical.However,they suffer from different shortages such as:higher fees,radioactivity and high false positive rate.Neither can be frequently used(more other than once per year)as a routine detection method.Microwave breast can-cer detection has high spatial resolution,non-invasive and lower cost which attracts more and more attention.It can be used as a potential routine detection methodology.The microwave-breast cancer detection can be roughly grouped into two directions:imaging and machine learning-based detection.Here we focus on the machine learning-based breast cancer detection.Empirical mode decomposition(EMD)and Dual-tree complex wavelet transformation(DTCWT)are firstly used to decompose the original high di-mensional waveform into different sub-bands,then some predefined statistical features are extracted from the sub-bands.This process reduce the dimension of original data.One ensemble classifier is built by using 2v-Support vector machine(SVM)as base classifier and Neyman-Pearson rules as criterion.The ensemble classifier can obtain the classifier whose empirical false positive is under the desired false positive rate.Additionally,in many practical applications,there is one common phe?nomenon that the number of samples collected from anomalous situation is far from the samples collected from the normal case,especially in medical and fi-nance fields.The ratio between healthy samples and malignant samples is close to 3:1 in our breast cancer dataset.This phenomenon is known as "Imbalanced learning" which may leads to a sub-optimal model for canonical machine learn-ing methods.Because the canonical machine learning learnings usually are de-signed by assuming that the ratio of samples of different classes are close to one.We also proposed one distribution-based data re-sampling method for im-balance learning in which the distribution in formation of each training sample were calculated.We selected the confidential subset of original dataset,accord-ing to the anomalous degree values which were estimated from both majority class set and minority class set.After that we generated synthetic minority class samples to rebalance the dataset.We finally proposed one hybrid-model for semi-supervised learning.The model combines deep auto encoder and nearest neighbour-based anomaly de-tector which is naturally suitable for processing the large-scale high dimen-sional data.The auto encoder was trained to transform the original dataset into a lower feature space.The samples in the lower feature space had a more com-pact distribution,compared to the original dataset.Thus,we could build the nearest neighbour-based anomaly detectors without losing the accuracy by us-ing the subset of the whole dataset.We then proposed one improved LPE-based semi-supervised anomaly detector which improved the robustness of distance estimation.The final decision was obtained by combining the outputs of all anomaly detectors.All the proposed algorithms were tested on our clinical breast cancer dataset and some open real-life datasets.The results confirmed that the proposed meth-ods improved the effectiveness of detection.