A deep learning approach to remove contrast from contrast-enhanced CT for proton dose calculation

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Proton dose calculation typically requires Non-Contrast Enhanced CT (NCECT), while Contrast Enhanced CT (CECT) is used for tumor delineation.
• Tissue motion between NCECT and CECT scans introduces dosimetry uncertainties, particularly for moving tumors in the thorax and abdomen.

Purpose of the Study:

• To develop and evaluate a deep learning approach for generating NCECT directly from CECT.
• To reduce CT simulation time, patient imaging dose, and motion-induced uncertainties in proton therapy.

Main Methods:

• A deep neural network was trained to convert 3D CECT images into contrast-removed NCECT image patches.
• The model was trained and tested using 8000 image patch pairs from 20 patients' abdominal CECT/NCECT data after deformable registration.
• Dosimetric impact was evaluated using clinical proton therapy plans and patient data.

Main Results:

• The deep learning model achieved high accuracy with a Cosine Similarity of 0.988 and MSE of 0.002.
• Generated NCECTs significantly reduced mean HU differences compared to CECT (∼4.72 vs. ∼64.52), a ~93% improvement.
• Proton dose calculations using generated NCECTs showed minimal changes in PTV (3.5%) and GTV (5.5%) V100% compared to CECT, but highlighted dose differences at distal beam paths.

Conclusions:

• A deep learning method effectively generates NCECT from CECT.
• This technique can mitigate uncertainties in proton dose calculation arising from inter-scan tissue motion, enhancing treatment planning accuracy.