Deep Learning for Automated Elective Lymph Node Level Segmentation for Head and Neck Cancer Radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate lymph node (LN) segmentation is crucial for defining target volumes in head-and-neck cancer (HNC) radiotherapy.
• Manual segmentation of LN levels is time-consuming and prone to variability.
• Automated segmentation methods are needed to improve efficiency and consistency.

Purpose of the Study:

• To evaluate deep learning approaches for segmenting individual lymph node levels (I-V) in HNC.
• To compare the performance of 3D patch-based UNets, multi-view (MV) voxel classification, and a sequential UNet+MV model.
• To assess the accuracy of automated segmentations for defining planning target volumes (PTVs).

Main Methods:

• Three deep learning models (UNet, MV, UNet+MV) were trained and validated using CT scans from 70 HNC patients.
• Five-fold cross-validation and ensemble learning were employed.
• Performance was measured using Dice Similarity Coefficients (DSC) for individual LN levels and combined PTVs.

Main Results:

• The UNet+MV model achieved the highest median DSC (0.82) for individual LN levels.
• UNet+MV significantly outperformed other methods (p < 0.0001), yielding DSC of 0.87 for combined levels I-V and 0.91 for PTVs.
• Accurate segmentation of individual LN levels I-V was achieved, with UNet+MV further refining results.

Conclusions:

• Ensemble UNets can accurately segment individual lymph node levels I-V in HNC.
• The UNet+MV approach offers a significant improvement in segmentation accuracy for HNC radiotherapy.
• Automated segmentation has the potential to reduce manual segmentation burden and variability.