Author: William N. Duggar, Amirhossein Eskorouchi, Haifeng Wang 👨🔬
Affiliation: Mississippi State University, University of Mississippi Medical Center 🌍
Purpose:
Extracapsular extension (ECE) in lymph nodes represents a critical prognostic factor in head and neck squamous cell carcinoma (HNSCC), bearing important implications for staging, treatment planning, and patient outcomes. Current methods for ECE detection rely on visual assessment of imaging or pathological confirmation, which can be subjective, time-intensive, and lack interpretability. This study presents a knowledge-informed deep learning (KIDL) framework that incorporates anatomical knowledge into CT imaging analysis to improve the accuracy and explainability of ECE identification.
Methods:
The KIDL framework employs a 3D DenseNet architecture with an informed decoder to generate binary masks that localize clinically relevant regions of interest, alongside a classification model to predict ECE status. A knowledge-regularized loss function was used to enforce consistency between model predictions and expert-defined anatomical features, reducing false positives. A total of 98 patients with confirmed HNSCC had their CT scans analyzed for this study, using a five-fold cross-validation protocol.
Results:
The model demonstrated an accuracy of 80%, an area under the receiver operating characteristic curve (AUC) of 83%, and a specificity of 88%. Importantly, the binary masks from the framework had overlapped regions with radiologists in regard to clinical significance; hence, this provides a visually interpretable output to help toward clinical decision making.
Conclusion:
This work demonstrates the potential of the KIDL framework in improving the clinical workflow for ECE detection in HNSCC by bringing together robust predictive accuracy with enhanced interpretability. Embedding domain-specific anatomical knowledge into the framework makes its output clinically meaningful and allows clinicians to make more consistent and efficient diagnoses. The approach bridges the gap between artificial intelligence and clinical practice and helps further the cause of precision oncology and personalized treatment planning.