A Logarithmic Mean Optimized EfficientNet with Joint Contrastive Learning for Improving Breast Cancer Detection from Mammogram Images

Abstract

Early prediction of breast cancer plays a critical role in improving patient survival rates and effective treatment planning. Mammography is the most often utilized imaging method for early breast cancer monitoring because of its ability to detect microcalcifications and subtle tissue abnormalities at an early stage. However, Mammography suffers from low sensitivity in dense breast tissue, overlapping structures and inter-observer variability, leading to high false positive rates and potential misdiagnosis. Recently, Deep Learning (DL) techniques have been utilised for mammography-based breast cancer detection, as they can automatically learn features from images. However, some models struggle with the conflation of malignant and benign areas, irrelevant feature extraction, and suboptimal training strategies. Their reliance on labelled data and classification-focused learning limits discriminative feature representation, resulting in poor differentiation of visually similar benign and malignant tissues. Hence, in this paper, an advanced DL approach is developed for efficient breast carcinoma prediction using mammography images. Initially, collected mammography images are augmented to enhance data diversity and reduce overfitting. Then, EfficientNet B7 with Joint Contrastive Learning (ENB7-JCL) is developed to learn discriminative visual representations. In ENB7-JCL, EfficientNetB7 is employed for deep feature extraction from mammography images, while JCL jointly optimizes contrastive and classification objectives by modelling similarity and dissimilarity relationships between samples. This improves the class separability and reduces misclassification of visually similar benign and malignant tissues. Additionally, the Logarithmic Mean Optimization Algorithm (LMOA) is applied to optimize the hyperparameters of EfficientNet B7 for effective training. Finally, a softmax classifier is used to perform breast cancer classification. The complete work is termed Optimized ENB7-JCL (OENB7-JCL). The experimental states that the proposed model achieves 94.85% and 95.12% of accuracy on the MIAS and INBreast dataset outperforming other existing models.