Effects of Image Augmentation and Dual-layer Transfer Machine Learning Architecture on Tumor Classification
Cheng Chen, Christine Chen, Xuesong Mei, Chaoyang Chen, Guoxin Ni, Stephen Edward Lemos · 2019
Breast tumor (BT) is the second most common health problem for women. Traditional diagnosis methods can be very labor-intensive and time-consuming with the risk of making a wrong diagnosis. Computer vision and imaging processing techniques using machine learning (ML) methods are emerging to aide in clinical diagnosis. Some machine learning methods have yielded an accuracy of 85% using a single-layer classifier. In this study Inception-V3, a two-layer classifier of transfer machine learning tool was used for image processing with enhancement technologies and for the classification of breast tumor histopathological types. Results showed that image augmentation with dual-layer transfer machine learning algorithms yielded an accuracy of 95.6% in identification of breast tumor pathologic types, which was higher than previously reported methods in the literature. Different image preprocessing methods, dataset preparing methods, and classifier architectures were also studied to identify the optimal algorithm. Results showed that multiple-layer processing algorithms using color images, instead of black and white images, yielded a better accuracy in histopathological type classification.