Optimized Mammography Preprocessing for Tumor Detection with YOLO11-Seg in Young Women from the Bolivian Altiplano

Abstract

Early diagnosis of breast cancer in young women presents a critical clinical challenge, particularly in geographic contexts such as the Bolivian Altiplano, where high breast density and limited access to specialized technologies hinder detection. This study evaluates the impact of various image preprocessing techniques on the performance of an automatic detection model based on YOLO11-seg. Using a dataset of mammograms annotated by certified radiologists, transformations such as CLAHE, histogram equalization, Canny filtering, wavelets, and anisotropic diffusion were applied. Standard metrics (mAP, precision, recall) were measured and results were compared in a real clinical setting. Findings show that CLAHE significantly improves the model's ability to detect lesions in dense breasts, achieving a mAP of 71.8%. The results suggest that combining enhancement techniques with AI models can strengthen early detection in high-risk populations, offering a viable and scalable alternative for resource-limited settings.