Automated detection of metastatic lymph nodes in head and neck malignant tumors on high - resolution MRI images using an improved convolutional neural network.

PURPOSE: To develop an AI-based diagnostic model for assessing cervical lymph nodes in head and neck malignant tumors using MRI, enabling non-invasive pre-surgical metastasis diagnosis.

MATERIALS AND

METHODS

Fifty-three cases of head and neck malignant tumors were retrospectively analyzed, including 157 metastatic lymph nodes and 2,406 MRI images. The dataset was split into training, validation, and test sets. A convolutional neural network (CNN) model was optimized through ablation and comparative experiments, and its diagnostic performance was evaluated using metrics such as average precision (AP), recall (AR), and mean average precision (mAP). A clinical evaluation compared the model's diagnostic efficiency to senior and junior physicians, assessing accuracy, sensitivity, specificity, predictive values, and area under the curve (AUC).

RESULTS

The model achieved detection and segmentation metrics of APdet 74.88 %, APseg 74.12 %, ARdet 63.11 %, ARseg 62.28 %, mAPdet 74.64 %, and mAPseg 74.04 %. Diagnostic accuracy was 83.6 %, with sensitivity 81.3 %, specificity 85.9 %, and an AUC of 0.834. The model processed the test set in 400 s (under 1 s per image), outperforming senior (AUC 0.706) and junior physicians (AUC 0.650), who required 1368 and 2276 s, respectively (p < 0.001).

CONCLUSION

The LNMS Net model enhances diagnostic accuracy and efficiency for head and neck malignant tumors, supporting precise treatment planning and reducing overtreatment risks. It also offers a foundation for extending AI-based lymph node metastasis diagnosis to other clinical areas.