| A New Classification Approach for Breast Cancer Classification based on Deep Leaning |
| Paper ID : 1120-ISCHU |
| Authors |
|
ASHRAF DARWISH1, MOURAD REEFAT2, Ahmed Labib3, Yehia Alaa *4 1Faculty of scince, Helwan university, Cairo,Egypt 2Faculty of scince,Helwan university, Cairo,Egypt 3Faculty of Science, Helwan University, Cairo, Egypt 4Mathematics Department, Faculty of Science, Helwan University, Cairo, Egypt. |
| Abstract |
| Cancer, particularly breast cancer, has become one of the most lethal diseases recently. Breast cancer was responsible for a very huge number of deaths worldwide recently, according to the World Health Organization (WHO). Chances of dying from breast cancer are significantly very high. Though, there has been a decreasing in breast cancer death rates due to the developing means of medical treatment, especially lately with using the increasingly popular developed AI techniques such as machine learning algorithms and deep neural networks to predict future outcomes. This paper describes a methodology used for classifying between calcified benign residuals and calcified malignant residuals that cause breast cancer tumors. This methodology was done using CNN with over 1.7 million trainable parameters along with means of augmentation. The dataset in regard was discovered to be insufficient, necessitating the use of this image augmentation technology which is interpolation-based image augmentation algorithms that returns the dataset to a usable form. The RBIS-DDSM dataset in question, that includes 849 images of breast cancer along with annotations for the calcified residuals have gone under computer vision analysis and it is revealed that only 475 images had either benign residuals or malignant residuals, while the remaining images either having both at once or having neither. Using interpolation to address the problem of an unbalanced dataset, a result of 865 images were gained that is better balanced and contains only one type of tumor per image. Consequently, fitting a built up CNN consists of 5 layers each is having an addition such as dropout layer, and maxpooling layer and a final flatten layer is added too, over the resulted dataset has led to classification with accuracy of the model 89.5% and the loss of data has reached 7%. Also, the F1_score for the first class and the second class was 0.88 and 0.91 respectively. |
| Keywords |
| AI - Machine Learning - Deep Learning - Convolutional Neural Network - CNN - Classification - Augmentation - Medical Images |
| Status: Abstract Accepted (Poster Presentation) |