# Mammogram Analysis with YOLO Models on an Affordable Embedded System

## TL;DR

This paper shows how affordable embedded systems can use YOLO models to detect breast cancer in mammograms, making early diagnosis accessible in low-resource areas.

## Contribution

The study introduces YOLOv11n as an optimal model for embedded mammogram analysis, enabling low-cost, high-accuracy lesion detection.

## Key findings

- YOLOv11n achieved 0.86 accuracy on NVIDIA Jetson Nano for mammographic lesion detection.
- The Jetson Nano offers comparable performance to GPU systems at a lower cost.
- A custom annotation tool improved labeling accuracy using a dataset of 3169 patients.

## Abstract

Breast cancer is a leading cause of death among women worldwide, and mammograms are a crucial tool for early detection. However, many resource-limited hospitals face challenges in accessing skilled radiologists and advanced diagnostic systems. This study evaluates the use of various You Only Look Once (YOLO) models, including YOLOv5, YOLOv8, YOLOv10, YOLOv11, and Real-Time-DEtection TRansformer (RT-DETR), for automated mammographic lesion detection on affordable embedded systems, such as the NVIDIA Jetson Nano (NVIDIA Corp., Santa Clara, California, USA). The results demonstrate that the YOLOv11n model performs optimally on this low-cost hardware, achieving an accuracy of 0.86 and an inference speed of 6.16 ± 0.31 frames per second. This research shows that deep learning-based computer-aided detection (CAD) systems can be deployed in low-resource clinical settings, improving access to early breast cancer detection in underserved regions.

Background/Objectives: Breast cancer persists as a leading cause of female mortality globally. Mammograms are a key screening tool for early detection, although many resource-limited hospitals lack access to skilled radiologists and advanced diagnostic tools. Deep learning-based computer-aided detection (CAD) systems can assist radiologists by automating lesion detection and classification. This study investigates the performance of various You Only Look Once (YOLO) models and a Hybrid Convolutional-Transformer Architecture (YOLOv5, YOLOv8, YOLOv10, YOLOv11, and Real-Time-DEtection Transformer (RT-DETR)) for detecting mammographic lesions on an affordable embedded system. Methods: We developed a custom web-based annotation tool to enhance mammogram labeling accuracy, using a dataset of 3169 patients from Thailand and expert annotations from three radiologists. Lesions were classified into six categories: Masses Benign (MB), Calcifications Benign (CB), Associated Features Benign (AFB), Masses Malignant (MM), Calcifications Malignant (CM), and Associated Features Malignant (AFM). Results: Our results show that the YOLOv11n model is the optimal choice for the NVIDIA Jetson Nano, achieving an accuracy of 0.86 and an inference speed of 6.16 ± 0.31 frames per second. A comparative analysis with a graphics processing unit (GPU)-powered system revealed that the Jetson Nano achieves comparable detection performance at a fraction of the cost. Conclusions: The current research landscape has not yet integrated advanced YOLO versions for embedded deployment in mammography. This method could facilitate screening in clinics without high-end workstations, demonstrating the feasibility of deploying CAD systems in low-resource environments and underscoring its potential for real-world clinical applications.