Deep Convolutional Neural Networks for Lesion Detection in Digital Mammograms

Abstract

Early detection of breast cancer significantly improves treatment outcomes and patient survival rates. This study proposes a deep learning framework for classifying benign and malignant breast lesions using the curated CBIS-DDSM dataset. The preprocessing pipeline includes three key steps: Gaussian filtering for noise reduction, CLAHE-based contrast enhancement to improve lesion visibility, and extensive data augmentation to promote generalization. We systematically evaluate three transfer learning-based CNN architectures (InceptionV3, EfficientNetB0, and MobileNetV2) initialized with ImageNet weights and fine-tuned for binary classification. Experimental results show that InceptionV3 achieves the best performance, reaching 90.2% accuracy, 87.1% sensitivity, 89.3% specificity, and an F1-score of 88.4% after only 70 epochs. EfficientNetB0 provides a more efficient alternative with 88.3% accuracy and faster convergence, while MobileNetV2, with 86.1% accuracy and 82.4% sensitivity, is well-suited for deployment in resource-constrained environments. These results demonstrate that a well-designed preprocessing strategy combined with transfer learning can yield clinically relevant classification performance, with InceptionV3 emerging as the most balanced and reliable architecture.