Deep learning model fusion improves lung tumor segmentation accuracy across variable training-to-test dataset ratios.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiotherapy Planning

Background:

• Accurate segmentation of Gross Tumor Volumes (GTVs) is crucial for lung cancer radiotherapy.
• Deep learning (DL) models show promise but can struggle with limited training data.
• Investigating model robustness at low training-to-test ratios (TTR) is essential for clinical translation.

Purpose of the Study:

• To evaluate the robustness of a DL fusion model for GTV segmentation in lung cancer SBRT.
• To assess performance across various low training-to-test ratios (TTR).
• To compare fusion models against individual DL models.

Main Methods:

• Developed and evaluated 12 DL models, including 3 individual models (3D U-Net, V-Net, dense V-Net) and 9 fusion models (AND, OR, voting).
• Utilized 192 lung cancer patient CT scans for training, validation, and testing at TTRs ranging from 1.00 to 0.116.
• Assessed segmentation accuracy using Dice Similarity Coefficients (DSC) and Hausdorff Distance (HD).

Main Results:

• The voting fusion model consistently achieved the highest DSCs (0.829–0.798) across all TTRs.
• Compared to single DL models, the voting fusion model demonstrated superior performance, especially at low TTRs.
• The voting fusion model showed favorable Hausdorff Distance (HD) values, indicating precise boundary delineation.

Conclusions:

• The proposed voting fusion model is a robust and effective method for segmenting lung cancer GTVs in planning CT images.
• This approach maintains high accuracy even with limited training data (low TTR), crucial for SBRT.
• The findings support the clinical utility of this DL fusion strategy in radiotherapy workflows.