How much data do you need? An analysis of pelvic multi-organ segmentation in a limited data context

Area of Science:

• Medical Imaging
• Deep Learning
• Radiotherapy

Background:

• Deep learning models require large datasets for training, limiting their use in specialized segmentation tasks.
• Pelvic multi-organ segmentation is crucial for radiotherapy planning.
• Limited data availability and domain specificity pose challenges for model generalizability.

Purpose of the Study:

• To evaluate the performance of the nnU-Net model for pelvic multi-organ segmentation under limited data conditions.
• To investigate the trade-off between dataset size and segmentation accuracy.
• To assess the impact of data augmentation on model performance with scarce data.

Main Methods:

• nnU-Net model trained on progressively smaller subsets of pelvic MR images (n=58 total, 46 for training, 12 for testing).
• Evaluation of segmentation performance using Dice Similarity Coefficient, mean surface distance, and 95% Hausdorff distance.
• Comparison of models trained with and without data augmentation.

Main Results:

• nnU-Net achieved high segmentation accuracy with the full training dataset (mean Dice: 0.903 Prostate, 0.851 SV, 0.884 Rectum, 0.967 Bladder).
• Performance remained stable with >12 training images but dropped significantly below this threshold.
• Data augmentation consistently improved performance across all dataset sizes, particularly for very small datasets.

Conclusions:

• nnU-Net demonstrates proficiency in male pelvic multi-organ segmentation under limited data and single-scanner constraints.
• A data threshold of approximately 12 images exists, below which performance degrades substantially.
• Data augmentation is a critical factor for maintaining performance in low-data scenarios, making nnU-Net valuable for in-house applications with limited data.