Deep learning-based computed tomography detection of early lymph node metastasis in head and neck cancer.

Background: Cervical lymph node metastasis (LNM) significantly influences the prognosis of patients with head and neck squamous cell carcinoma (HNSCC). However, conventional computed tomography (CT) diagnostics are susceptible to high false-negative rates for small lesions, exhibit considerable inter-reader variability, and are labor-intensive due to the requirement for manual evaluation by radiologists. This study aimed to develop and validate a deep learning-based model incorporating an attention mechanism, termed the deep learning-based cervical lymph node metastasis detection model (DL-CervLNM), to automatically detect LNM in contrast-enhanced CT scans. This model was designed to address challenges related to the detection of small lesions, inter-observer variability, and diagnostic efficiency. Methods: A total of 6,860 contrast-enhanced CT scans from 481 patients with HNSCC at Fujian Cancer Hospital between 2020 and 2024 were retrospectively collected and analyzed for model development. An asymmetric context-aware cascade detection (ACCD) framework was introduced, consisting of two principal modules: (I) an attention-enhanced You Only Look Once version 8 (YOLOv8) stage, specifically designed to generate candidate regions with high recall rates; and (II) a region-based refinement stage that incorporates the Faster region-based convolutional neural network (R-CNN) algorithm to enable context-aware suppression of false positives. The efficacy of the ACCD framework was rigorously assessed in comparison to board-certified radiologists and state-of-the-art baseline models. The primary evaluation metrics included the mean average precision (mAP) at an intersection-over-union (IoU) threshold of 0.5 (mAP@0.5), the area under the receiver operating characteristic curve (AUC), the F1-score, and computational efficiency metrics. Results: In the context of LNM detection using CT imaging, DL-CervLNM achieved a mAP@0.5 of 94.9% and an AUC of 0.980, outperforming both current state-of-the-art models and experienced radiologists. The model exhibited enhanced sensitivity (90.6%) and specificity (94.5%), with a notable proficiency in identifying small lesions, achieving an accuracy of 93.2% compared to 76.4% for radiologists. In addition, the system improved clinical workflow efficiency by enabling real-time processing at 38.5 frames per second (FPS), thereby significantly reducing interpretation and reporting times by 51.7% and 98.6%, respectively. Conclusions: DL-CervLNM exhibited enhanced accuracy and reliability in detecting cervical LNM on CT scans compared to current methodologies. This model has the potential to improve diagnostic efficiency and consistency, thereby aiding radiologists in the early identification of metastatic lymph nodes.