Clinical Scorecard: Addressing Class Imbalance in Deep Learning for Segmentation of Head and Neck Organs

At a Glance

Key Highlights

• Manual segmentation of HAN organs is time-consuming, subjective, and causes treatment delays.
• Deep learning segmentation is challenged by class imbalance from varying organ sizes and background.
• Optimizing training patch size and adapting Dice loss improves segmentation performance and confidence.

Guideline-Based Recommendations

Diagnosis

• Use CT imaging for planning and identification of organs at risk in HAN cancer radiotherapy.

Management

• Apply deep learning-based automated segmentation to reduce manual workload and observer variability.
• Optimize training patch size to balance class representation during network training.
• Incorporate class adaptive Dice loss to handle missing classes within training patches.
• Consider cascaded networks segmenting large organs first, then small organs to reduce class imbalance.
• Combine cross-entropy and Dice loss functions to improve segmentation accuracy.

Monitoring & Follow-up

• Evaluate segmentation performance using Dice score (DSC) and Hausdorff distance metrics.
• Perform multi-class confidence analysis to assess segmentation reliability across organ sizes.

Risks

• Class imbalance may lead to overfitting of small organ classes and reduced segmentation accuracy.
• Intrinsic bias of Dice loss towards large volumes can impair small organ segmentation.

Patient & Prescribing Data

Patients with head and neck cancers requiring radiotherapy planning

Automated segmentation using optimized deep learning approaches can reduce treatment delays and improve accuracy of organ delineation.

Clinical Best Practices

• Train segmentation networks with patch sizes optimized to minimize class imbalance.
• Use class adaptive Dice loss to improve robustness against missing classes in patches.
• Employ cascaded or hybrid 2D/3D convolutional network architectures to address class imbalance.
• Combine voxel-wise cross-entropy loss with Dice loss for balanced training.
• Analyze segmentation confidence to detect potential overfitting in small organ classes.

References

• Cancer statistics and HAN tumor incidence
• Observer variations in manual segmentation
• Deep learning for HAN organ segmentation
• First DL-based multi-organ HAN segmentation approach
• nnU-Net framework for segmentation
• Multi-class confidence analysis for class imbalance
• Cascaded networks addressing class imbalance
• Small organ segmentation with autoencoder priors and adversarial training
• FocusNetv2 for HAN segmentation
• Autoencoder-based shape priors
• Adversarial training in segmentation
• MICCAI 2015 HAN segmentation challenge dataset
• Hybrid 2D/3D convolutional networks for HAN segmentation
• UNet architecture
• Class weighted cross-entropy loss
• Dice loss for segmentation
• Generalized Dice score
• Combination of Dice and focal loss