The InterVision Framework: An Enhanced Fine-Tuning Deep Learning Strategy for Auto-Segmentation in Head and Neck

Area of Science:

• Medical Physics
• Radiotherapy Oncology
• Artificial Intelligence in Medicine

Background:

• Adaptive radiotherapy (ART) aims for precise tumor targeting and organ sparing by adapting treatment plans to anatomical changes.
• Current ART workflows face challenges in real-time implementation due to manual recontouring and time constraints.
• Deep learning-based segmentation (DLS) shows promise for automating contouring but requires extensive, high-quality datasets for generalizability.

Purpose of the Study:

• To develop and evaluate the InterVision framework for improving deep learning-based segmentation accuracy in adaptive radiotherapy.
• To address the challenge of limited high-quality datasets for DLS in clinical ART settings.
• To enhance the generalizability and performance of segmentation models by incorporating patient-specific characteristics.

Main Methods:

• The InterVision framework generates intermediate images between existing patient scans using deformable vectors to capture unique anatomical variations.
• A two-step training process was employed: first, a general model was trained on the dataset, followed by fine-tuning using data generated by the InterVision framework.
• Segmentation models were evaluated using the volumetric dice similarity coefficient (VDSC) and Hausdorff distance 95% (HD95) for 18 structures in 20 patients.

Main Results:

• The InterVision model achieved a higher Dice score (0.85 ± 0.03) compared to the general model (0.81 ± 0.05) and the general fine-tuning model (0.82 ± 0.04).
• The InterVision model demonstrated a lower Hausdorff distance (2.52 ± 0.50) compared to the general model (3.06 ± 1.13) and the general fine-tuning model (2.81 ± 0.77).
• The InterVision framework significantly improved segmentation accuracy, particularly for complex organs and targets.

Conclusions:

• The InterVision framework offers a versatile and effective approach to enhance DLS for ART by creating more comprehensive, patient-specific datasets.
• This method addresses the critical need for accurate and efficient auto-segmentation in dynamic radiotherapy environments.
• The InterVision framework shows potential for broader applications in medical imaging where prior information is available.