Improving Breast Cancer Diagnosis in Mammograms with Progressive Transfer Learning and Ensemble Deep Learning

This paper presents a novel approach that integrates ensemble deep learning and progressive transfer learning to enhance breast cancer diagnosis in mammographic images, leveraging the INbreast dataset. The study commences with an exhaustive examination of two ensemble learning techniques, bagging and stacking, and introduces a novel concept, progressive transfer stacking. Harnessing the capabilities of deep learning and transfer learning, our method unfolds gradually, streamlining the diagnostic into three interconnected steps, each managed by a dedicated convolutional neural network (CNN). Five distinct CNN architectures-VGG16, ResNet, EfficientNet, DenseNet, and Xception-are progressively trained across the three diagnostic steps. The initial step categorizes mammogram images into three classes: Normal, Benign/Malignant, or Muscle. Subsequent steps refine the classification, differentiating Benign/Malignant images into Benign or Malignant and further classifying them into six distinct subclasses. Notably, the dataset exhibits non-balanced classes, a challenge effectively addressed through ensemble learning, specifically via stacking numerous CNN models. This strategic application of ensemble learning significantly contributes to overcoming imbalanced class distributions within the dataset. The multi-step diagnostic process markedly enhances precision, culminating in a noteworthy F1-score of 99% in the final classification of twelve different breast cancer classes. Importantly, our approach stands out as the first to address the classification of 12 classes, not just a binary classification of malignant or benign, marking a substantial advancement in the field of medical image classification.