Author: Liyuan Chen, Sepeadeh Radpour, David Sher, Jing Wang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center 🌍
Purpose: Accurate lymph node malignancy prediction is pivotal in optimizing radiation treatment strategies for head and neck (HN) cancer patients. While conventional radiomics models leverage intensity, geometric, and texture-based features, they often overlook critical spatial and anatomical factors associated with lymphatic dissemination. This study introduces a novel spatially aware radiomics model that integrates lymphatic system anatomy and clinical knowledge to improve the performance for lymph node malignancy prediction.
Methods: A dataset comprising 1,320 lymph nodes (1,091 benign and 229 malignant) with diameters smaller than 17 mm, contoured on pre-operative CT scans from 192 HN cancer patients, was utilized to train and evaluate the proposed model. Two radiomics models were developed: a baseline model using conventional features and an enhanced model incorporating five additional features. These features, informed by lymphatic anatomy and clinical knowledge, included the distance between the lymph node and primary tumor, primary tumor type, cervical level of the lymph node, and laterality of both the lymph node and the primary tumor. Model performance was assessed using sensitivity (Sen), specificity (Spe), accuracy (ACC), and the area under the receiver operating characteristic curve (AUC).
Results: The baseline model achieved Sen = 0.915, Spe = 0.756, ACC = 0.787, and AUC = 0.931. The enhanced model, incorporating spatial and anatomical features, demonstrated significant improvements with Sen = 0.919, Spe = 0.845, ACC = 0.860, and AUC = 0.953.
Conclusion: This study demonstrates the value of incorporating lymphatic system anatomy and clinical knowledge into radiomics models to improve lymph node malignancy prediction. By incorporating prior clinical knowledge, the performance of the baseline radiomics model utilizing conventional features was significantly improved.