Deep Learning-Based Detection of Malignant and Equivocal Cervical Lymph Nodes on CT Imaging Using a 2D U-Net++ Model.

Introduction: Accurate identification of malignant and equivocal lymph nodes on computed tomography (CT) remains a significant challenge in clinical practice, particularly in borderline cases. Recent advances in artificial intelligence offer potential solutions for improving diagnostic performance and supporting radiological decision-making. Materials and methods: A dataset of 79 CT examinations acquired between January 2024 and December 2025 was retrospectively analyzed, including 538 annotated lymph nodes classified as Node-RADS 3-5. A 2D U-Net++ convolutional neural network was implemented and trained in a slice-wise manner using NIfTI-formatted data. The dataset was divided into training and validation subsets, with data augmentation being applied through random rotations. Model performance was evaluated using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), sensitivity, specificity, precision and accuracy. An independent test set of 12 CT cases was used for qualitative assessment. Results: The model achieved an accuracy of 99.8% on the training set and 99.7% on validation, with specificity consistently reaching 99.9%. Sensitivity was 78% for training and 72% for validation, while precision reached 83% and 79%, respectively. Segmentation performance yielded a DSC of 0.79 (training) and 0.73 (validation), with IoU values of 0.67 and 0.61. The model demonstrated robust anatomical discrimination, correctly excluding common mimics such as accessory spleens, salivary glands, vascular structures and muscular tissue. Qualitative evaluation showed limited false-negative results (three equivocal lymph nodes) and two false-positive detections corresponding to benign (Node-RADS 1) lymph nodes. Conclusions: The proposed 2D U-Net++ model demonstrates reliable performance in the detection of malignant and equivocal lymph nodes on CT imaging, with high specificity and good anatomical discrimination. Its applicability to both contrast-enhanced and non-contrast CT scans supports its potential as a clinical decision-support tool, particularly for highlighting borderline findings in routine practice.