Breast Cancer is a leading cause of women’s mortality worldwide. Therefore, accurate and early mammary cancer diagnosis with mammograms can improve disease detection accuracy at an early stage. To assist and enhance the ability of radiologists in the proper detection of breast lesions, digital mammograms must be preprocessed, enhanced, segmented, and classified by computer-aided diagnosis systems (CAD). The high rate of false positives and false negatives in the detection process is a major cause of low survival rates for women. An early or well-timed diagnosis can help to improve the woman’s chances of healing and recovery. To this end, we propose an optimized classification framework based on transfer learning and enhanced deep convolutional neural networks (DCNNs) to accurately categorize mammographic images as benign or malignant. To ensure consistency in the preprocessing phase we adopted a previously established segmentation method for pectoral muscle removal, as detailed in [1], then a modified contrast enhancement method is proposed to refine the visibility, detail, and the edges of the mammary tissues to make the representation of abnormalities or tissue with lesions clear based on Singular Decomposition Value (SVD) and Discrete Wavelet Transform (DWT). For the classification stage, a fine-tuned version of MobileNetV2, along with comparative analysis using InceptionResNet-V2, NasNetLarge, Xception is implemented to evaluate the performance on MIAS Dataset in order to improve classification and categorization performance. The proposed method achieved an average accuracy of 96.88% demonstrating a strong classification performance. These findings highlight the benefit of incorporating refined preprocessing with advanced transfer learning techniques, ultimately contributing to more reliable mammogram interpretation and reducing false positive and false negative rates in clinical screening.