Impact of multisequence MRI on deep learning-based dose prediction for glioblastoma radiotherapy: A comparative evaluation of CT-only and CT+MRI models.

INTRODUCTION Deep learning-based dose prediction has emerged as a promising approach for supporting automated radiotherapy planning workflows by estimating three-dimensional dose distributions from patient imaging. Although planning CT remains the standard imaging modality, multisequence MRI provides anatomical and tissue-characterization information that may improve prediction accuracy in glioblastoma (GBM) radiotherapy. This study investigated the incremental value of multisequence MRI by comparing CT-only and CT + MRI models using an identical framework.

Methods

A retrospective cohort of 140 patients with GBM treated with conventionally fractionated radiotherapy (60 Gy in 30 fractions) was analyzed. Two 3D Attention U-Net models were developed: CT-only and CT + MRI incorporating T1-weighted gadolinium-enhanced (T1GD), FLAIR, and T2∗ MRI sequences. Patients were divided into training (n = 100), validation (n = 20), and test (n = 20) cohorts. Models were trained for 130 epochs using Adam optimizer with a composite loss function of 70% mean absolute error (MAE) and 30% mean squared error.

Results

The CT + MRI model achieved lower MAE than CT-only (5.99 ± 1.45 Gy vs. 6.92 ± 1.74 Gy, p = 0.006). D2 error decreased from 3.51 ± 1.61 Gy to 1.53 ± 1.53 Gy. SSIM improved from 0.564 ± 0.102 to 0.599 ± 0.134 (p = 0.017), and PSNR increased from 16.84 ± 2.17 dB to 18.01 ± 1.92 dB (p = 0.009). Pearson correlation coefficients were 0.929 and 0.937 for CT-only and CT + MRI models, respectively.

Conclusion

Multisequence MRI improved deep learning-based dose prediction for GBM radiotherapy, particularly spatial dose agreement and high-dose region estimation. Further multicentre validation is required. IMPLICATIONS FOR PRACTICE MRI-enhanced dose prediction may support automated planning workflows by improving predicted dose distributions.