Author: Liyuan Chen, Meixu Chen, Bowen Jing, Sepeadeh Radpour, Erich Josef Schmitz, David Sher, Jing Wang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Prospective clinical trials have shown that involved nodal radiation therapy (INRT) can substantially improve patients’ quality of life without increasing the risk of elective nodal failure. In INRT, only lymph nodes (LNs) classified as malignant will receive a therapeutic dose, as opposed to the whole neck region in conventional radiotherapy. Therefore, accurate identification of malignant LNs is critical for the success of INRT. Previous LN malignancy prediction models have only focused on individual nodes without considering the presence of malignant LNs in nearby regions or within the same nodal station that can profoundly affect malignancy status. To model connections between nearby LNs, we developed a graph-based neural network that treats a patient’s LNs as a graph, aggregating image features from connected LNs to improve prediction accuracy.
Methods: Preoperative CT images from 192 patients were used, totaling 1,391 LNs with pathological ground truth. CT images were divided into patches centered on individual LNs which were then used as vertices in a graph structure. The graphs were used as input to a deep neural network comprising a convolutional neural network (CNN) and a graph neural network (GNN). The malignancy probability was predicted using features aggregated from connected LNs. Performance was evaluated using nested 5-fold cross validation and compared to a baseline network without the GNN component.
Results: The CNN+GNN network outperformed the CNN-only model in predicting LN malignancy in HNC. The CNN+GNN model achieved a Receiver Operating Characteristic Area Under the Curve (AUC), average precision (AP), sensitivity, and specificity of 0.940, 0.933, 0.880, and 0.879, respectively, while the CNN-only model achieved 0.921, 0.909, 0.859 and 0.861 for the corresponding evaluation criteria.
Conclusion: By incorporating the LNs of HNC patients into a graph, our model uses contextual information from surrounding LNs to enhance malignancy prediction for suspicious nodes.