Comparison of Vendor-Pretrained and Custom-Trained Deep Learning Segmentation Models for Head-and-Neck, Breast, and...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Radiotherapy planning

Background:

• Commercial deep learning (DL) models for medical image segmentation often show variable performance.
• Institutional data and clinical practices can influence the effectiveness of these DL segmentation tools.

Purpose of the Study:

• To evaluate the impact of local patient and clinical characteristics on the performance of commercial deep learning segmentation models.
• To compare the performance of vendor-pretrained versus custom-trained DL models for head-and-neck, breast, and prostate cancers.

Main Methods:

• Utilized clinical CT scans and contours from 210 patients across four cancer types.
• Trained and validated four custom DL segmentation models using institutional data.
• Assessed performance of both vendor-pretrained and custom-trained models using Dice Similarity Coefficient (DSC) and Mean Surface Distance (MSD).

Main Results:

• Custom-trained models demonstrated superior performance over vendor-pretrained models for 14 out of 24 organs at risk.
• Significant improvements in DSC and MSD were observed for breast and prostate cancer segmentation with custom models.
• Average DSC for custom models: HN (0.86), breast (0.80), prostate (0.92) vs. vendor models: HN (0.81), breast (0.67), prostate (0.87).

Conclusions:

• Institutional data variations significantly impact the performance of vendor-pretrained DL segmentation models.
• Customizing DL models with even limited institutional data can achieve sufficient accuracy for clinical applications.
• Tailoring DL segmentation models is crucial for optimizing their implementation in radiotherapy planning.