TRI-BCC: Tri-Level Breast Cancer Classification via Transfer Learning Networks with Histopathological Images

Abstract Breast cancer (BC) is the second most common hereditary cancer in women and is most likely to affect women. The high cost of treating BC has made its prevention an important health issue. However, early detection of BC is a time-consuming and difficult task when using mammogram images. This work introduces a novel TRI-BCC model for BC detection using transfer learning-based tri-level classification with histopathological images. Initially, input images undergo a denoising phase using the Adaptive Brightness Contrast Dynamic Histogram Equalization (ABCDE) technique, which removes noisy artifacts to enhance overall image quality. Following this pre-processing step, a data augmentation phase is employed to synthetically expand the training dataset, thereby improving model generalization and robustness against overfitting. The augmented and pre-processed images are subsequently provided as input to the hybrid Golden Eagle-Whale Optimization (GWO) algorithm, for the segmentation of lesions. In the tri-level classification, transfer learning (TL) models: CapsuleNet, EfficientNet, ShuffleNet, GoogleNet, MobileNet and ResNet are employed to categorize images as benign or malignant (Level-I). Benign cases (Level-II) include Adenosis (AS), Phyllodes tumor (PT), Tubular adenoma (TA) and Fibroadenoma (FA), while malignant cases (Level-II) include Papillary (PC), Mucinous (MC), Lobular (LC), and Ductal (DC) carcinoma. Malignant images are processed using a Randomized Decision Tree (RDT) to classify them into different cancer stages (Level-III) for precise diagnosis. Based on experimental analysis, the TRI-BCC model achieves an overall accuracy of 99.06 % for the categorization of BC. The proposed TRI-BCC model improves overall accuracy by 6.11 %, 0.78 %, 2.66 %, and 13.18 % compared to UNet+YOLO, fine-tuned networks, hybrid deep neural network, and Pa-DBN-BC, respectively.