Creative Commons source reading · Prediction and forecasting
Deep learning–based non-invasive prediction of axillary lymph node metastasis in breast cancer: performance of the YOLO-v11 object detection algorithm
Sevgi Ünal; Remzi Gürfidan; Merve Gürsoy Bulut; Oğuzhan Kilim
Originally published in BMC Med Imaging · 2026-03-16
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
BackgroundThis study proposes a deep learning-based decision support system for the non-invasive prediction of axillary lymph node (ALN) metastasis in breast cancer using ultrasound (USG). Both localisation and benign–malignant classification of ALNs were performed using YOLO-v11, a single-stage object detection approach. The dataset created from labelled USG images was split into training/validation/test sets at a ratio of 70%/20%/10%, and the model produced a class label and confidence score for each prediction. Performance was found to be high with amAP@0.5 = 0.904s (benign: 0.866, malignant: 0.942); F1-, precision-, and sensitivity-confidence curves showed that thresholds in the 0.52–0.60 range provided a balance between sensitivity and specificity. Confidence scores for the malignant class were observed to be concentrated in the…
Results
The four graphs shown in Fig. 2 show from different angles how the model behaves as the “confidence threshold” changes and provide a scientific basis for the choice of threshold. Fig. 2Threshold screening and performance curves by class (benign/malignant) The F1-Confidence Curve gives the threshold-dependent profile of the F1 score, which is the harmonic mean of precision and recall. Since F1 penalises both false positives and false negatives at the same time, it is used to determine the best “general purpose” threshold. In the graph, F1 for all classes plateaus broadly in the range 0.30–0.65 and peaks at F1 ≈ 0.86 at a threshold of approximately 0.52–0.55. This indicates that there is flexibility in the choice of threshold, but that F1 starts to fall rapidly above 0.70. Complementary to this, the Precision-Confidence Curve shows how…
Discussion
Today, axillary lymph node dissection and sentinel lymph node biopsy (SLNB) are accepted as standard surgical methods for axillary lymph node staging [18]. However, both operations carry varying degrees of risk for postoperative complications and morbidity. Recent studies have focused on minimising unnecessary axillary surgical interventions and avoiding overtreatment in breast cancer [18]. The Sentinel Node and Observation after Axillary Ultrasound (SOUND) trial demonstrated that in patients with tumor size ≤ 2 cm and negative findings on axillary ultrasonography, omission of SLNB is safe, provided that axillary staging is performed preoperatively with USG [19]. USG is the most commonly preferred imaging method in ALN assessment. However, the diagnostic accuracy of ultrasound largely depends on the radiologist’s experience [20].…
Editorial reading guide
- What data created and tested the model, and was evaluation independent?
- What is the cost of a false positive, false negative or out-of-distribution decision?
