Deep Learning and Registration-Based Mapping for Analyzing the Distribution of Nodal Metastases in Head and Neck...

Area of Science:

• Oncology
• Medical Imaging
• Radiotherapy

Background:

• Nodal metastases (LNs) are critical in head and neck (H/N) cancer staging and treatment.
• Accurate spatial analysis of LNs is essential for designing effective radiotherapy (RT) target volumes.
• Current methods for LN analysis can be labor-intensive and subjective.

Purpose of the Study:

• To introduce and evaluate deep learning and registration-based methods for automated spatial analysis of LNs in H/N cancer.
• To inform radiotherapy (RT) target volume design using these novel analytical tools.
• To compare the performance of the proposed methods against expert ground truth.

Main Methods:

• A deep learning approach utilizing an nnU-Net 3D/2D ensemble model for autosegmentation of 20 H/N levels.
• A nonrigid registration-based mapping method to analyze LN distribution in a cohort-average template CT.
• Kernel density estimation to determine the 3D-LN probability distribution without predefined levels.
• Multireader assessment by three radio-oncologists to establish ground truth for the deep learning method.

Main Results:

• The deep learning method achieved 100% accuracy in categorizing 449 LNs across 193 H/N patients.
• Level 2 demonstrated the highest nodal involvement, accounting for 59.0% of analyzed LNs.
• The registration-based mapping technique showed results consistent with the ground-truth distribution (p=0.915).

Conclusions:

• Automated deep learning and registration methods offer a robust approach for analyzing LN spatial distribution in H/N cancer.
• These methods can significantly aid in optimizing radiotherapy (RT) target volume design.
• Further application to multicenter cohorts and specific H/N tumor subtypes holds promise for advancing clinical practice.