Patient-specific deep learning model to enhance 4D-CBCT image for radiomics analysis

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Four-dimensional computed tomography (4D-CBCT) provides crucial phase-resolved imaging for radiomics analysis in radiotherapy outcome prediction.
• Artifacts in 4D-CBCT, stemming from under-sampling, degrade radiomic feature accuracy.
• Previous group-patient-trained models improved 4D-CBCT quality but lacked individual patient optimization.

Purpose of the Study:

• To develop and evaluate a patient-specific deep learning model for enhancing 4D-CBCT image quality.
• To improve the accuracy of radiomic features extracted from 4D-CBCT for individual patients.
• To assess the effectiveness of the patient-specific model compared to group-based models in radiotherapy.

Main Methods:

• A patient-specific model was trained using intra-patient data, augmenting 4D-CT scans to simulate patient positioning variations.
• Simulated 4D-CBCT data were reconstructed and paired with corresponding 4D-CT images for model training.
• The model was tested on four lung-SBRT patients, comparing radiomic features from original 4D-CT, 4D-CBCT, and enhanced 4D-CBCT.

Main Results:

• The patient-specific model significantly improved radiomic feature accuracy compared to group-based models, particularly for features with high error rates.
• Average error reductions for specific features included 83.67% for first-order median, 91.98% for wavelet LLL maximum, and 15.0% for wavelet HLL skewness.
• Model dimensionality (2D vs. 3D) and loss functions (L1, L1+VGG+GAN) showed comparable improvements; whole-body or whole-body+ROI L1 loss outperformed ROI L1 alone.

Conclusions:

• Patient-specific models are more effective than group-based models for enhancing 4D-CBCT radiomic feature accuracy.
• This approach holds potential for improving the precision of outcome prediction in radiotherapy.
• Further investigation into model dimensionality and loss functions can optimize performance.