What type of study is this?

This is a Human Clinical Trial study.

September 10, 2025Open Access

Radiomics Meets Deep Learning: A Hybrid Approach for Breast Cancer Prediction from Mammographic Data

Key Points

The hybrid model achieves high accuracy in predicting breast cancer from mammographic data, indicating effective early detection.
Data augmentation methods like rotation and scaling enhanced model generalization, showing significant impact on accuracy and performance.
Five CNN architectures, including VGG16 and DenseNet121, were evaluated, with custom models yielding promising results.
Deep learning approaches like this can improve medical image analysis, suggesting scalable solutions for breast cancer diagnostics.

Abstract

Breast cancer persists as the most commonly occurring malignancy in the female population around the world. Early and accurate detection is essential to enhance patient prognosis and the detection of pulmonary nodules using deep learning algorithms has recently been highlighted in medical image analysis. This work investigates hands-on deep neural learning approach for breast mammographic cancer prediction in early stages. Histopathology data are pre-processed (normalized and standard resized) to keeps the input size consistent and to maximize the features. A CNN architecture was constructed and trained using the Keras library, with TensorFlow as the backend. Data augmentation was used to regularization of the algo and to prevent high variance to the small amount of dataset, using rotation, flip, and scale manipulation. Models Five types of two-dimensional networks were developed: a custom CNN, VGG16 with frozen convolutional layers, and three networks (ResNet50, DenseNet121, and EfficientNetB3) trained from scratch. Of these, the custom CNN architecture showed potential: a high accuracy to discriminate between malignant and benign tumors. The accuracy and loss curves, as well as confusion matrices, confirmed the quality and consistency of the results. Interestingly, even relatively deep networks (such as DenseNet121 or EfficientNetB3) obtained better generalization in combination with data augmentation, also when trained from scratch. These results highlight the strength of deep convolutional models for medical images classification problems and suggest that the depth of the architecture and the right choice of regularization strategy are fundamental in achieving robust detection of cancer. This research further enriches the emerging area of AI-based diagnostics by providing comparative statistics for DL approaches in predicting the occurrence of breast cancer. Our results provide useful guidance for the future development of more accurate, scalable, and immediately practical predictors.

Radiomics Meets Deep Learning: A Hybrid Approach for Breast Cancer Prediction from Mammographic Data

Key Points

Abstract

Cite This Study