A Transfer Learning Framework for Deep Learning-Based CT-to-Perfusion Mapping on Lung Cancer Patients

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Deep learning models can estimate lung perfusion from CT scans.
• Performance of CT-based perfusion mapping in lung cancer patients requires specific investigation.

Purpose of the Study:

• To develop and evaluate a transfer learning framework for CT-to-perfusion mapping in lung cancer patients.
• To adapt existing deep learning models for accurate perfusion assessment in this specific patient group.

Main Methods:

• A transfer learning framework was created using SPECT/CT perfusion scans from 33 lung cancer and 137 non-cancer patients.
• A pre-trained model was fine-tuned on lung cancer data using three-fold cross-validation.
• Evaluation included voxel-wise correlation (Spearman's R), function-wise similarity (Dice Similarity Coefficient), and lobe-wise perfusion values.

Main Results:

• The fine-tuned model achieved a high voxel-wise correlation of 0.8142 ± 0.0669.
• Function-wise similarity showed average Dice Similarity Coefficients of 0.8112 ± 0.0484 (high-functional) and 0.8137 ± 0.0414 (low-functional).
• High Dice Similarity Coefficients (>0.7) were achieved for most functional lung volumes in lung cancer patients.

Conclusions:

• The transfer learning framework enables accurate CT-based lung perfusion mapping in lung cancer patients.
• The method shows strong correlation and similarity with SPECT/CT, validating its clinical potential.
• This approach holds promise for providing regional functional information crucial for lung-sparing radiation therapy planning.