Cervical lymph node image analysis using YOLO11 segmentation model: an application and performance study

Objectives Ultrasound assessment of superficial cervical lymph nodes is clinically valuable, yet current diagnostic methods are highly operator-dependent, with limited objectivity and consistency. This study investigates a more objective approach for analyzing lymph node ultrasound images, focusing on the stability and practical applicability of different model architectures on real-world clinical data. It aims to evaluate the utility of deep learning models for detection and segmentation of superficial lymph nodes and to provide a feasible methodology for objective lymph node ultrasound analysis. Methods Using real-world clinical data, a cervical lymph node ultrasound image dataset comprising 1,050 patients was constructed, with histopathological confirmation and standardized clinical-imaging follow-up (≥50 % diameter reduction over 6–9 months after treatment) serving as reference standards. Three different models – transformer-based RF-DETR, two-stage YOLO11+MedSAM, and end-to-end YOLO11-Seg – were systematically compared on detection and segmentation tasks within the same dataset. Results YOLO11-Seg demonstrated superior performance in both detection and segmentation. Under COCO evaluation metrics, it achieved an mean average precision (mAP) at 0.5 of 0.795 and an mAP at 0.5:0.95 of 0.558. Pixel-level segmentation evaluation yielded a Dice coefficient of 0.775 ± 0.15 and an IoU of 0.645 ± 0.15, with Precision and Recall values of 0.796 and 0.783, respectively. In contrast, the two-stage YOLO11+MedSAM and Transformer-based models exhibited lower accuracy and reduced consistency. Conclusions This study provides preliminary evidence for the application of deep learning–based methods in lymph node ultrasound analysis and suggests a potential technical framework for generating structured imaging data, which may support future computer-assisted assessment pending further validation.