Breast cancer detection using deep learning techniques: An investigation using the CBIS-DDSM dataset and customized neural network model, ResNetV2 and YOLO

Abstract

The current work investigates the application of deep learning methodologies for detecting and classifying tumors in mammograms. A comprehensive analysis was conducted by evaluating various pre-trained neural network architectures to identify the model that delivers optimal performance. Among the tested architectures, Residual Neural Networks and Densely Connected Convolutional Networks were explored extensively. Data preprocessing, including data augmentation, was a critical step due to the limited availability of public medical imaging datasets. This process ensured diversity in the data and improved model robustness. This paper evaluates deep-learning models for mammogram tumor detection. Experiments without the AdamW optimizer and horizontal flip showed overfitting and low precision (below 40%). Densely Connected Convolutional Networks achieved high precision but exhibited overfitting with noisy validation loss curves. Adding AdamW and horizontal flip reduced overfitting but lowered overall performance. In classification, the model detected tumors in 9 out of 16 images, showing potential but requiring improvement. The model’s ability to detect calcification tumors enhances robustness. YOLO network metrics were modest, reflecting the task’s complexity, but results were acceptable for tumor classification challenges.