Cancer Region Segmentation in Pre-Processed Breast Ultrasound Image Using VGG16 Based UNet/SegNet
Ramya Mohan, Mathiyazhagan Narayanan, V. Rajinikanth
Abstract
Breast Cancer (BC) is a predominant cause of mortality among women globally, with its incidence progressively rising due to various causes. Early detection of BC is crucial to plan and execute appropriate treatment to cure the disease. The clinical level examination of the BC involves in; personal check by the doctor, screening using non-invasive medical imaging modality, and biopsy test to confirm the BC and its stage. Compared to the radiological imaging modalities, the breast-ultrasound (BU) based screening is widely preferred due to its simplicity. The abnormality recorded with BU is then examined to classify the tumor as benign/malignant, an essential process during BC detection. This research aims to propose a deep-learning (DL) tool to segment the tumor from the chosen BU-image. To enhance the segmentation accuracy, every image is preprocessed using tri-level thresholding with Tsallis's Entropy and Hummingbird-Algorithm (TE+HA). The performance of proposed UNet/SegNet is verified on both unprocessed and preprocessed images. The experimental outcome of this research confirms that the VGG-UNet offered an improved mean accuracy (>97<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>) when the preprocessed BU is considered for examination.