Improving breast cancer prediction via progressive ensemble and image enhancement

Breast cancer, the most commonly diagnosed cancer in women worldwide. In areas with limited budgets, training qualified medical professionals to accurately diagnose breast cancer remains a challenge, particularly in the interpretation of mammogram images due to the subtle distinctions between benign and malignant lesions. While breast cancer patients need to be diagnosed as early as possible to increase the chance of cure. All these reasons raises a significant need for a more economical, timely and accurate solution. We introduce a novel combination of image enhancement techniques, including Gamma Correction, Contrast Limited Adaptive Histogram Equalization (CLAHE), Retinex, and Image Super-Resolution (ISR) - tailored to overcome these interpretive challenges by significantly improving image quality and detail visibility. Furthermore, we leverage progressive image resizing, an innovative technique that systematically increases the resolution of images during the model training process, to effectively capture detailed patterns in mammogram evaluation. Additionally, we present a fine-tuning strategy for pre-trained models such as ResNet-50, EfficientNet-B5, and Xception, combining multiple preprocessing methods and extracting inherent features through transfer learning to improve model reliability and classification accuracy. Finally, we systematically compare three ensemble methods: averaging, voting, and weighted averaging, with the latter showing superior accuracy for breast cancer detection classification results. This approach synergizes each model’s distinct feature extraction strengths, culminating in a high predictive performance. Progressive image resizing from 150 \(\times \) $\times$ 150 to 240 \(\times \) $\times$ 240 improves model generalization. Ensemble modeling by averaging, voting, and weighted averaging predictions achieves up to 91.36% accuracy for mass/calcification classification and 76.79% for benign/malignant classification. This study develops an accurate deep learning framework for breast cancer prediction that holds promise to assist radiologists and improve patient care, utilizing the publicly accessible Curated Breast Imaging Subset of the Digital Database for Screening Mammography dataset (CBIS-DDSM).