# Spatially aware radiomics integrating anatomical knowledge to improve lymph node malignancy prediction in head and neck cancer

## TL;DR

This study improves cancer diagnosis by adding anatomical and spatial information to radiomic models, leading to better prediction of malignant lymph nodes in head and neck cancer.

## Contribution

The novel spatially aware radiomics model integrates anatomical knowledge and clinical factors to enhance malignancy prediction in lymph nodes.

## Key findings

- The enhanced model achieved higher accuracy (ACC = 0.860) and AUC (0.953) compared to the baseline model.
- Incorporating spatial and anatomical features significantly improved model performance (p = 3.71 × 10−20 for accuracy).

## Abstract

Radiomics holds the potential to improve the diagnostic evaluation of equivocal lymph nodes in head and neck cancer (HNC). While conventional radiomics models utilize features such as intensity, geometry, and texture of individual lymph node, they often neglect key spatial and anatomical characteristics tied to lymphatic dissemination patterns.

In this study, we propose a novel spatially aware radiomics model that integrates anatomical knowledge and clinical factors to enhance lymph node malignancy prediction.

A total of 1389 lymph nodes (1119 benign and 270 malignant), contoured on CT scans from 192 HNC patients were included. Two models were developed: a baseline model using conventional radiomics features and an enhanced model incorporating five additional spatial and anatomical features, such as primary tumor type, lymph node level, the laterality of the primary tumor, the laterality of the lymph node, and the distance from the lymph node to the primary tumor. Sensitivity (SEN), specificity (SPE), accuracy (ACC), positive predictive value (PPV), negative predictive value (NPV) and the area under the receiver operating characteristic curve (AUC) criteria were used to evaluate the model performance.

The proposed spatially aware radiomics model significantly outperformed the baseline model. The baseline model achieved SEN = 0.915, SPE = 0.756, ACC = 0.787, PPV = 0.475, NPV = 0.974, and AUC = 0.931. The enhanced model achieved SEN = 0.919, SPE = 0.845, ACC = 0.860, PPV = 0.589, NPV = 0.977, and AUC = 0.953. Statistical testing confirmed a significant improvement in both accuracy (p = 3.71 × 10−20) and AUC (p = 1.13 × 10−4).

This study demonstrates that incorporating lymphatic anatomy and clinical context into radiomics models significantly improves predictive performance. The proposed approach enhances interpretability, aligns with clinical workflows, and holds promises for personalized radiation therapy planning.