LARGEST ROI SEGMENTATION FOR BREAST CANCER CLASSIFICATION USING A VGG16 DEEP LEARNING NETWORK
Thanh-Tam NGUYEN, Thanh-Hai NGUYEN, Ba-Viet NGO, Thanh-Nghia NGUYEN
DOI: 10.15598/aeee.v22i4.240303
Abstract
The exact evaluation of breast cancer images for patients is very important, because they can be early treated for lasting their life. This article proposes a classification system for finding breast cancer images, in which each breast lesion image is segmented to produce a largest Region of Interest (ROI) and a VGG16 deep learning network is applied for classification. An Otsu threshold is utilized on two datasets from two sources of CBIS-DDSM and MIAS to create largest ROI with main features. For the classification with high performance, two datasets of the breast lesions were augmented by rotating, flipping, and brightness variation. This article was proposed an algorithm with processing images sets before classification using VGG16. In particular, the results of the largest ROI datasets for four types of breast lesions were represented through segmentation, normalization and enhancement. In addition, the results of classifying four types of breast lesions (BC, BM, MC, MM) were evaluated using confusion matrix, with the high accuracy of around 95%. Another evaluation was that these image sets without ROI/with ROI parts/With the largest ROI only using the Otsu segmentation were compared and the highest accuracy was of the image sets with the largest ROI. The results with the high accuracy demonstrated to illustrate the effectiveness of the proposed method. It means that this method can be developed to classify many stages of breast cancers during diagnosis and treatment.