Purpose: To develop a machine learning-based, 3D dose prediction methodology for Gamma Knife treatment plans.
Methods: We obtained 322 Gamma Knife (GK) treatment plans from our institution, under REB approval. To handle the sparsity, size, and granularity of GK data, several data augmentations were made. The contoured MRI and clinical dose distributions were isolated and cropped based on tumor location and scaled to a standard size through interpolation. An accompanying 3D tensor containing distance information was created for each instance of cropped data to account for tumor size. This augmented dataset for 272 patients was used to train a convolution neural network (CNN-GK) and a generative adversarial network (GAN-GK). As a baseline, the same CNN (CNN-Baseline) and GAN (GAN-Baseline) were trained on the same dataset without the GK-specific data augmentations described above. All models were used to predict the dose distribution of 50 out-of-sample patients. Predictions were compared to their clinical counterparts using the gamma criterion (4%/2mm) and conformity index.
Results: The predictions from CNN-GK and GAN-GK were generally similar to the clinical dose distributions, with gamma passing rates of 83.1 ± 17.2% and 84.9 ± 15.3%, respectively. In contrast, the gamma passing rate of CNN-Baseline (36.2 ± 33.8%) and GAN-Baseline (40.3 ± 36.4%) were both significantly lower than their respective GK-specific models (p < 0.001). Additionally, for 41% of the tumours in the out-of-sample patients, both baseline models predicted dose with a gamma pass rate of 0.0%. The conformity index of CNN-GK and GAN-GK predictions differed, on average, from that of the clinical distributions by 0.092 ± 0.10 and 0.086 ± 0.11 respectively.
Conclusion: Our methodology that uses deep learning models and GK-specific data augmentation can accurately predict 3D dose distributions for Gamma Knife treatment plans. Off-the-shelf deep learning models applied to standard GK data generally generate poor predictions.