Dose Distribution Prediction for Hepatocellular Carcinoma Using Convolutional Neural Networks from Diagnostic CT and MRI: Focus on Passive and Intensity-Modulated Proton Therapy

Background Proton therapy is commonly used for hepatocellular carcinoma (HCC). However, its feasibility can be challenging to assess large tumors or those adjacent to critical organs at risk (OARs), as these factors are typically evaluated only after treatment planning. This study aimed to predict proton dose distributions utilizing diagnostic computed tomography (dCT) and magnetic resonance (MR) images, leveraging a convolutional neural network (CNN), to enable early treatment feasibility assessments before planning CT (pCT) acquisition. Methods This study calculated dose distributions and dose-volume histograms (DVH) for 118 patients with HCC using intensity-modulated proton therapy (IMPT) and passive proton irradiation. The CNN model predicted the DVH and dose distributions, which were evaluated using mean absolute error (MAE), mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Results The predicted DVHs closely matched the actual DVHs. MAE was consistently below 0.03, with passive proton therapy achieving values between 0.012 and 0.018, indicating high consistency. MSE remained below 0.004 for all cases, confirming clinically acceptable accuracy. PSNR ranged from 24 to 28 dB. SSIM was above 0.94 for both IMPT and passive proton therapy, with the lowest value being 0.841, indicating high structural similarity. Conclusions This study demonstrates the potential of diagnostic imaging in optimizing the workflow for HCC proton therapy planning. The proposed CNN-based model enables early dose distribution predictions, reducing the need for pCT acquisition and improving treatment decision-making.