Axillary lymph node (ALN) metastasis is one of the strongest determinants of breast cancer prognosis and directly influences decisions about surgical approach, systemic therapy, and postoperative radiotherapy. Approximately 20 to 30% of breast cancer cases develop axillary metastasis, making accurate preoperative assessment a fundamental component of staging and treatment planning.
Ultrasonography (USG) is the primary imaging method for evaluating axillary lymph nodes before surgery. Radiologists use criteria such as lymph node size, shape, cortical thickness, and loss of the fatty hilum to identify suspicious nodes. However, USG has a critical limitation: it is highly operator-dependent, with accuracy varying substantially based on the radiologist's experience, equipment settings, and patient anatomy.
A further challenge is that approximately 35% of metastatic axillary lymph nodes show no suspicious features on ultrasound at all, making them visually indistinguishable from normal nodes. This combination of operator dependency and false-negative rate motivates the development of AI tools that could provide automated, reproducible lymph node classification to assist radiologists in real clinical settings.
This study developed a deep learning model based on YOLO-v11 (You Only Look Once, version 11) to simultaneously localize and classify axillary lymph nodes in breast ultrasound images as either benign or malignant. Unlike classification-only models that require the lymph node to be pre-identified, YOLO-based object detection performs both tasks in a single pass, outputting a bounding box around each detected lymph node and a confidence score for its class label.
The dataset comprised 471 patients from a single institution in Turkey: 248 with biopsy-confirmed malignant axillary lymph nodes and 223 control cases with benign nodes. This represents the first application of a YOLO-based architecture specifically to axillary lymph node classification in breast cancer -- prior YOLO work in this domain focused on other anatomical sites or different cancer types.
A key methodological feature was patient-level splitting of the dataset into training (70%), validation (20%), and test (10%) sets. All images from a single patient were assigned exclusively to one subset, ensuring that the test set contained only truly unseen patients. This approach prevents data leakage, where similar images from the same patient in both training and test sets could inflate performance estimates.
The YOLO (You Only Look Once) family of architectures is designed for single-stage object detection: instead of first proposing candidate regions and then classifying them (as two-stage detectors like Faster R-CNN do), YOLO processes the entire image in one forward pass, simultaneously predicting bounding box coordinates and class probabilities. This single-stage approach makes YOLO architectures faster and more computationally efficient than two-stage methods, while maintaining competitive accuracy.
Images were resized to 960x960 pixels using YOLO's letterbox strategy, which preserves aspect ratio and applies padding to prevent geometric distortion. Pixel intensities were normalized to a 0-1 range. No additional preprocessing such as contrast enhancement, denoising, or histogram equalization was applied -- the model was trained directly on clinically typical raw ultrasound images, an important choice that makes the pipeline applicable in real clinical workflows without additional processing steps.
Data augmentation used the default Ultralytics YOLO pipeline, applying geometric transformations and scale-based augmentations to increase effective training diversity. The model was trained over 200 epochs with three parallel loss components: box regression loss for bounding box accuracy, classification loss for benign/malignant discrimination, and Distribution Focal Loss (DFL) for refined box boundary prediction. Training and validation loss curves showed steady convergence without signs of overfitting.
The model achieved a mean Average Precision at IoU 0.5 (mAP@0.5) of 0.904 overall, with 0.866 for benign and 0.942 for malignant lymph nodes. The mAP@0.5 metric measures the area under the precision-recall curve at a localization overlap threshold of 50%, combining both detection accuracy and classification quality into a single number. A value of 0.90+ indicates strong and reliable detection performance.
The malignant class consistently outperformed the benign class across all metrics. Confidence scores for malignant detections were concentrated in a tight band of 0.82 to 0.87, indicating that the model assigned consistently high confidence to its malignant predictions. Benign confidence scores were somewhat more dispersed in the 0.70 to 0.80 range. The authors attribute the malignant class advantage to the more distinctive visual features of metastatic nodes: cortical thickening, irregular margins, and heterogeneous hypoechoic internal structure -- all strong discriminative cues for deep feature extraction.
On the independent test set (32 images, 37 annotated lymph nodes), the normalized confusion matrix showed benign accuracy of approximately 0.89 and malignant accuracy of 0.84. The main source of errors was confusion with the image background: 31 detections did not sufficiently overlap with any annotated lymph node, representing false positives caused by vascular cross-sections, fascial planes, or other anatomical structures with node-like ultrasound appearance.
Because different clinical scenarios require different tradeoffs between sensitivity and specificity, the study analyzed how model performance varies with the confidence threshold used to decide whether a detection is accepted. The F1-confidence curve showed a plateau between thresholds of approximately 0.30 and 0.65, with a peak F1 score of approximately 0.86 at a threshold of 0.52 to 0.55, indicating that the model is relatively robust to threshold choice within this range.
For screening scenarios where missing a malignant node is clinically dangerous, the recall-confidence curve shows that sensitivity remains above 0.9 at low thresholds but drops rapidly above 0.65 to 0.75. This means a radiologist prioritizing sensitivity should operate at a threshold of around 0.52 to 0.60, accepting some false positives in exchange for very high malignancy detection rates.
For high-specificity scenarios where false positives must be minimized, the precision-confidence curve shows that precision approaches 1.0 at thresholds around 0.77 to 0.78, meaning nearly all predictions at this level are truly malignant. The practical clinical guidance the authors provide is to use the 0.52 to 0.60 confidence band for general-purpose balanced use, and to increase to 0.75 to 0.80 when false positive cost is particularly high.
Qualitative visualization of model detections revealed consistent patterns in how YOLO-v11 distinguished benign from malignant nodes. For benign lymph nodes, the model placed bounding boxes that followed the long axis of the lesion horizontally, encompassing the smooth contours and homogeneous echo texture characteristic of normal nodes. Confidence scores ranged from 0.70 to 0.80, reflecting appropriate uncertainty about subtle benign cases.
For malignant lymph nodes, boxes tightly surrounded foci with irregular contours, with the heterogeneous hypoechoic nucleus more prominently captured. Confidence scores in the 0.82 to 0.87 range were consistent across malignant cases, suggesting the model reliably detected the morphological fingerprint of metastatic involvement. The model also showed contextual awareness, intentionally preserving surrounding tissue context such as stromal striation and fascia transitions that provide reinforcing classification cues.
The model was also able to detect and separately annotate multiple lymph nodes within a single ultrasound frame -- a clinically important capability since patients often have several axillary nodes visible simultaneously. However, this also introduced challenges: in benign cases, the model sometimes produced multiple overlapping bounding boxes for weakly echogenic structures or similar-appearing tissue regions, suggesting the Non-Maximum Suppression (NMS) threshold allowed some redundant benign detections through.
Prior AI studies on ALN ultrasound classification have used a range of architectures. Ozaki et al. using the Xception classification network reported 94% sensitivity, 88% specificity, and AUC 0.966 using classification-only CNN approaches. Coronado-Gutierrez et al. achieved 86.4% accuracy, 84.9% sensitivity, and 87.7% specificity on 118 lymph node images. Wang et al. using YOLO-v7 on cervical lymph nodes reported mAP of 96.4% with accuracy values above 0.96.
The YOLO-v11 results in this study (benign accuracy 0.89, malignant accuracy 0.84, mAP@0.5 of 0.904) are broadly comparable to prior classification-only approaches and somewhat below the best cervical node YOLO-v7 results. The authors note that differences in disease region (axillary vs. cervical), data diversity, and image quality make direct comparison difficult. The key contribution is the simultaneous localization and classification in a single forward pass -- a practical advantage over classification-only systems that require prior lymph node identification.
A primary source of false positives in this study was confusion between lymph nodes and background anatomical structures including vascular cross-sections and fascial planes. These errors were concentrated in small bounding boxes at image peripheries and at low confidence thresholds. The linear, repetitive texture of ultrasound tissue also contributed to background false positives, particularly when context from surrounding tissue was lost in very small detections.
The study positions the YOLO-v11 model as a decision support tool rather than a replacement for radiologist evaluation. In the context of the SOUND trial results -- which demonstrated that SLNB omission is safe in patients with tumors up to 2 cm and negative axillary ultrasound -- an AI tool that improves the accuracy and reproducibility of axillary ultrasound assessment becomes directly clinically actionable in expanding SLNB-avoidance decisions.
The real-time, automated nature of YOLO-based detection is a practical advantage for clinical deployment. A YOLO model can analyze ultrasound images during the scan itself, flagging suspicious lymph nodes immediately and providing a confidence score that the radiologist can accept, modify, or override. This contrasts with approaches that require separate preprocessing, manual region-of-interest marking, or batch analysis after imaging is completed.
The model was trained on clinically typical raw ultrasound images without preprocessing, and the study deliberately preserved clinical image annotations (patient information, scale bars, lesion markers) during training to reflect real-world acquisition conditions. This design choice increases the likelihood that performance estimates reflect what would be achieved in actual clinical use, rather than in idealized pre-processed data conditions.
This study demonstrated that YOLO-v11 can simultaneously localize and classify axillary lymph nodes in breast ultrasound images with clinically meaningful accuracy (mAP@0.5 = 0.904), representing the first application of this architecture to this specific task. The model's higher malignant sensitivity and consistent confidence scores in the 0.82 to 0.87 range for malignant detections are aligned with the clinical priority of minimizing missed cancers.
Key limitations include the single-center design using data from two ultrasound systems at the same institution, which limits assessment of generalizability across different hospitals, operators, and equipment manufacturers. The independent test set of only 32 images is too small to provide statistically precise performance estimates, though confidence intervals were estimated via bootstrap resampling to quantify this uncertainty. No direct comparison to radiologist performance was performed.
Future directions include multicentre external validation across different institutions, ultrasound manufacturers, and clinical environments; cross-device domain adaptation strategies to address equipment-specific variability; and prospective evaluation comparing AI-assisted to unassisted radiologist performance. These steps are essential before the system can be considered for clinical deployment as a validated decision support tool.