Learning Curves for Auto-contouring of Head and Neck Tumors Using CT & PET Scans

Abstract

Medical image segmentation is one of the key applications of artificial intelligence in the medical field. With the widespread adoption of deep learning methods in medical image segmentation, researchers are increasingly exploring their potential for segmentation of head and neck tumors. This study focuses on the automatic head and neck tumor segmentation based on CT and PET, and investigates the performance variation of models under different training sample sizes and data source combination strategies.

The study employs the 3D U-Net model, by utilizing the open dataset HECKTOR 2022 and the closed dataset LUMC, three training and testing configurations were designed: training and testing on the same dataset, direct cross-dataset testing, and hybrid training incorporating a small amount of target closed dataset. By increasing the number of training samples, the learning curve characteristics of the model were examined utilizing metrics.

Results indicate that under consistent training and testing data sources, increasing the training dataset size significantly enhances model segmentation performance and stability. However, beyond certain data sizes, performance improvements tend to flatten out. In cross-dataset evaluation, merely expanding the source dataset's training scale yields limited gains for target dataset performance. Conversely, incorporating a small number of training samples consistent with the testing data source substantially improves model performance and learning curve stability.

Overall, this study clarifies the effects of training dataset size and data source consistency on automated head and neck tumor segmentation from a learning curve perspective, providing empirical evidence for model performance in cross-dataset application scenarios.