Systematic Evaluation of Pre-Processing and Post-Processing Strategies for Head-and-Neck Organ Segmentation
Abstract
Purpose
To evaluate and optimize segmentation pipeline configurations for head-and-neck organ segmentation to improve accuracy, robustness, and clinical applicability.
Methods
We systematically evaluated the impact of preprocessing, spacing configuration, and post-processing strategies on head-and-neck organ segmentation. An automated anatomy-guided ROI cropping pipeline was developed to localize the head-and-neck region without manual intervention. Body and bone were used to estimate cropping boundaries based on HU values, and the resulting bounding box was applied consistently to CT images and segmentation labels. ROI cropping was applied to reduce irrelevant background and improve training efficiency and segmentation accuracy. Three voxel spacing configurations were evaluated to assess the effect of spatial resolution on segmentation performance. Post-processing techniques were used to suppress false-positive predictions. Experiments were conducted on an internal head-and-neck dataset with 400 patients using the nnU-Net framework, and performance was evaluated using the Dice similarity coefficient (DSC).
Results
Experimental results showed that ROI cropping provided the largest performance improvement. Under isotropic spacing (1x1x1 mm3) with a 128x128x128 patch size, ROI cropping increased the average DSC from 0.766 to 0.78, demonstrating the effect of irrelevant background exclusion. ROI cropping significantly accelerated convergence from 849 to 211 epochs. Spacing and patch configuration further improved segmentation performance. The default nnU-Net spacing (1.27x1.27x3 mm3) achieved an average DSC of 0.775, while spacing (1x1x2 mm3) with a 160x160x80 patch size achieved an average DSC of 0.777. Post-processing provided an additional performance gain by removing false-positive predictions, improving the average DSC from 0.777 to 0.780.
Conclusion
This study demonstrates that segmentation pipeline configuration significantly affects head-and-neck organ segmentation performance. The optimization of spacing, patch size, and ROI cropping is important for accurate head-and-neck organ segmentation.