Classification of Mammography Images by Machine Learning Techniques

Bektas B., Entre I. E., Kartal E., Gulsecen S.

3rd International Conference on Computer Science and Engineering (UBMK), Sarajevo, Bosna-Hersek, 20 - 23 Eylül 2018, ss.580-585 identifier identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Cilt numarası:
  • Doi Numarası: 10.1109/ubmk.2018.8566380
  • Basıldığı Şehir: Sarajevo
  • Basıldığı Ülke: Bosna-Hersek
  • Sayfa Sayıları: ss.580-585
  • Anahtar Kelimeler: machine learning, breast cancer diagnosis, feature extraction, feature selection, classification
  • İstanbul Üniversitesi Adresli: Evet

Özet

Early diagno.sis and accurate treatment is crucial in increasing the survival rate of diseases that can result in death, such as breast cancer. Therefore, there is a greater need for artificial intelligence systems that will help doctors make decisions in health care, especially in fatal diseases. Because these systems are not affected by human nature factors such as distraction, stress etc. so that they can distinguish small and important issues that could be overlooked, especially in the scan results of the patient. The aim of this study is to predict whether a mass can be identified in breast and whether the mass found in the breast is benign or malignant with the help of machine learning which is a sub study area of artificial intelligence. In this study, the images in the mini-MIAS database are used. Firstly, unwanted areas were eliminated. Then Gauss, Average, Median and Wiener filters were applied to reduce noise and smoothing the images and an algorithm based on Contrast-Limited Adaptive Histogram Equalization (CLAHE) was applied to make suspicious areas more visible. New data sets were created by using HOG (Histogram of Oriented Gradients), LBP (Local Binary Pattern), LLCM (Gray Level Co-occurrence Matrices) for feature extraction and correlation (COR) for feature selection. Selected features were classified in three different categories (normal, benign, malignant) and two different categories (normal, abnormal) using. Different machine learning algorithms (C5.0 (normal and boosted), Naive Bayes, CART and Random Forest) were applied to the data sets and the performances were compared. According to the research findings, to decide whether there was a breast mass, the highest accuracy value was calculated by applying Median and Wiener filters together, equating histogram with CLAHE and using the GLCM feature extraction method on the data set and the accuracy was found 0.657 with Naive Bayes algorithm. When trying to find out whether the mass found in the breast is benign or malignant, Median was applied together with Weiner Filter, equating histogram with CLARE and HOG feature extraction method was used, and the accuracy was calculated as 0,660 with Random Forest algorithm.