Addressing Class Imbalance in Deep Learning for Head and Neck Organ Segmentation

Overview

Class imbalance significantly impacts the performance of deep learning models for segmenting head and neck (HAN) organs, especially small soft tissue structures. This work optimizes training patch size and introduces a class adaptive Dice loss to improve segmentation accuracy and confidence, particularly for mid-sized and small organs.

Background

Head and neck cancers represent about 3% of newly diagnosed cancers and require precise image-guided radiotherapy. Manual segmentation of organs at risk (OAR) on CT scans is time-consuming and subject to observer variability. Deep learning (DL) approaches have become the dominant method for automated segmentation but face challenges due to class imbalance caused by large size differences among organs and between foreground and background voxels. Addressing this imbalance is critical to improving segmentation performance across all organ sizes.

Data Highlights

The FocusNetv2 approach achieved a Dice score (DSC) of 0.84 and a 95% Hausdorff distance (95HD) of 2.17 mm on the MICCAI 2015 HAN segmentation challenge dataset, representing the current best reported results. The proposed method incorporating class imbalance optimized patches and class adaptive Dice loss improves upon the baseline nnU-Net framework, increasing segmentation confidence and accuracy, especially for mid-sized organs.

Key Findings

• Training patch size directly influences class imbalance; optimizing patch size reduces imbalance and improves segmentation performance.
• A class adaptive Dice loss formulation robustly handles missing classes within patches, enhancing training stability with sparse class distributions.
• Incorporating these methods into the nnU-Net framework increases segmentation accuracy beyond the baseline model.
• Multi-class confidence analysis reveals increased segmentation confidence for mid-sized organs when using class imbalance optimized patches.
• Previous state-of-the-art methods addressed class imbalance via cascaded networks and hybrid 2D/3D architectures, but this work focuses on patch size and loss function adaptations.
• Combining Dice loss with other loss functions like focal loss or cross-entropy has shown improvements, but intrinsic biases toward large volumes remain a challenge.

Clinical Implications

Optimizing training patch size and employing a class adaptive Dice loss can enhance automated segmentation of head and neck organs, reducing manual workload and variability. Improved segmentation accuracy, especially for small and mid-sized organs, can facilitate more precise radiotherapy planning and potentially reduce treatment delays.

Conclusion

Addressing class imbalance through patch size optimization and adaptive loss functions significantly improves deep learning-based segmentation of head and neck organs. These advancements support more reliable and efficient clinical workflows in radiotherapy planning.

References

1. MICCAI 2015 HAN Segmentation Challenge Dataset
2. Isensee et al. 2021 -- nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation
3. Gao et al. 2021 -- FocusNetv2 for Head and Neck Organ Segmentation
4. Li et al. 2020 -- Multi-class Confidence Analysis in Imbalanced Segmentation
5. Milletari et al. 2016 -- V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation