Enhancing Mammography Models: The Impact of Radiologist Recommendations on Algorithmic Precision

Abstract

This study demonstrates the substantial benefits of integrating advanced image classification techniques into the diagnosis and treatment of breast cancer. Our comprehensive approach utilizes deep learning algorithms, with a focus on enhancing the reliability and efficiency of mammography image classification. Specifically, we employ YOLOv5 for precise image segmentation and Densenet121 for extracting informative features from segmented regions of interest (ROIs). The dataset, comprising 54,706 mammography images, facilitates both visual and numerical analyses. To optimize breast cancer diagnosis, we strategically examine a separate set of 100 challenging cases, randomly selecting 50 positive and 50 negative instances, ensuring a balanced representation of both benign and malignant cases. Careful validation involves curating 50 consensus cases for reliability. To address class imbalance, we employ a hybrid Upsampling/Downsampling approach. Moreover, we fine-tune 14 algorithms, comparing outcomes with and without radiologists' recommendations. The analysis, particularly focusing on the AUC (Area Under the Curve), reveals compelling results. Without radiologists' recommendations, the model achieves a 99.8\% AUC during testing and 59.5\% during validation. Incorporating radiologists' insights notably boosts AUC to 99.9\% for testing and 93.5\% for validation. These findings emphasize the pivotal role of expert guidance in refining diagnostic accuracy. This study delves into the interplay between algorithmic precision, dataset characteristics, and expert recommendations in breast cancer diagnosis, providing valuable insights for leveraging advanced technologies and expert knowledge for improved medical outcomes.