Deep Learning-Based Cardiac Chamber Segmentation in Magnetic Resonance-Guided Adaptive Radiation Therapy

Area of Science:

• Medical Imaging
• Radiation Oncology
• Cardiovascular Imaging

Background:

• Accurate cardiac chamber segmentation is vital for cardiac sparing in magnetic resonance (MR)-guided adaptive radiation therapy.
• Patients undergoing radiation therapy are at risk for radiation-induced cardiotoxicity, necessitating precise targeting to minimize heart exposure.

Purpose of the Study:

• To develop and evaluate automatic segmentation models for cardiac chambers using daily MR images from a 1.5-T MR-Linac.
• To assess the performance of segmentation models based on T2/T1 3DVaneXD balanced fast field echo with spectral attenuated inversion recovery (bFFE-SPAIR) and T1 3DVaneXD mDixon sequences.

Main Methods:

• Three 3D nnU-Net models were trained: bFFE-SPAIR (bFFE model), T1 mDixon (mDixon model), and a hybrid model.
• Models were evaluated using Dice Similarity Coefficient (DSC) and mean surface distance against manual contours.
• Clinical acceptance was assessed via a 5-point Likert scale, and an in-silico study evaluated cardiac chamber sparing during adaptive planning.

Main Results:

• The bFFE model achieved the highest segmentation performance (average DSC 0.85 ± 0.05).
• The T1 mDixon sequence provided similar accuracy (DSC 0.83 ± 0.06), despite lower contrast-to-noise ratios.
• 95% of bFFE model autosegmented contours were clinically acceptable (score ≥4), and adaptive planning significantly reduced cardiac dose.

Conclusions:

• Feasibility of accurate cardiac chamber segmentation on a 1.5-T MR-Linac using bFFE-SPAIR and T1 mDixon sequences is demonstrated.
• Developed automatic segmentation models show potential for enhancing cardiac sparing in MR-guided adaptive radiation therapy.
• These advancements can contribute to improved patient outcomes by minimizing radiation-induced cardiotoxicity.