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

Abstract

This study highlights the benefits of advanced image classification in breast cancer diagnosis and treatment. We utilize deep learning algorithms like YOLOv5 for image segmentation and Densenet121 for feature extraction from segmented regions. Our dataset includes 54,706 mammography images for comprehensive analysis. We evaluate 100 challenging cases, ensuring a balanced representation of benign and malignant instances. Validation involves 50 consensus cases. To address the class imbalance, we employ Upsampling/Downsampling. We fine-tune 14 algorithms and compare outcomes with and without radiologists' recommendations. Results show a 99.8\% AUC during testing and 59.5\% during validation without radiologists' input, which improves to 99.9\% and 93.5\% respectively with their insights. Expert guidance significantly enhances diagnostic accuracy. The study explores the interplay between algorithmic precision, dataset characteristics, and expert recommendations in breast cancer diagnosis. It provides valuable insights for leveraging technology and expert knowledge for improved medical outcomes.