Tackling the class imbalance problem of deep learning-based head and neck organ segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Accurate segmentation of organs at risk (OAR) is crucial for image-guided radiation therapy.
• Deep learning (DL) models face challenges in segmenting head and neck (HAN) areas due to class imbalance and anatomical size variations.
• Automating OAR segmentation is vital for efficient and effective cancer treatment.

Purpose of the Study:

• To address the class imbalance issue in HAN OAR segmentation using DL.
• To optimize the nnU-Net framework for improved segmentation performance in imbalanced datasets.
• To introduce and evaluate a class adaptive Dice loss function for enhanced segmentation accuracy.

Main Methods:

• Optimized patch size for the nnU-Net framework based on a novel class imbalance measurement.
• Introduced a class adaptive Dice loss function to mitigate the effects of severe class imbalance.
• Applied these methods to segment seven organs in the head and neck region.

Main Results:

• Optimized patch size and class adaptive Dice loss resulted in a 3% increase in Dice score.
• Achieved a 22% reduction in the 95% Hausdorff distance compared to baseline methods.
• Demonstrated high segmentation accuracy for seven HAN organs, reaching [Formula: see text] and [Formula: see text] mm.

Conclusions:

• Patch size optimization and class adaptive Dice loss are effective strategies for improving DL segmentation in imbalanced scenarios.
• These methods are easily integrated into existing DL segmentation frameworks.
• The study highlights a simple yet effective approach to enhance OAR segmentation for radiation therapy.