Generalization vs. Specificity: In Which Cases Should a Clinic Train its Own Segmentation Models?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Artificial intelligence (AI) for medical image segmentation is rapidly advancing.
• Generalizability of AI models across different institutions and geographic locations remains a critical question.

Purpose of the Study:

• To compare the performance of multi-institutional AI models against site-specific models for organs-at-risk segmentation.
• To differentiate the impact of population variability versus clinical practice differences on AI model generalizability.
• To provide guidelines for selecting between custom-trained and third-party AI segmentation models.

Main Methods:

• Utilized CT datasets from four clinics covering female breast, female pelvis, and male pelvis organs-at-risk.
• Compared segmentation quality between multi-institutional and site-specific deep neural networks.
• Analyzed the influence of training set size and patient population variability.

Main Results:

• Segmentation quality for female pelvis organs and the heart depended primarily on training set size.
• Patient population variability significantly impacted female breast segmentation quality, exceeding the effect of training set size.
• For the male pelvis, site-specific models outperformed multi-institutional ones with large datasets, but multi-institutional models were superior with small datasets.

Conclusions:

• The choice between multi-institutional and site-specific AI models depends on dataset size and anatomical region.
• Site-specific models are advantageous for large datasets, particularly for male pelvis organs.
• Multi-institutional models offer better performance for smaller datasets, especially for organs like the female pelvis and heart.