Dose distribution prediction in isodose feature‐preserving voxelization domain using deep convolutional neural network

Abstract

Purpose: To implement a framework for dose prediction using a deep convolutional neural network (CNN) based on the concept of isodose feature-preserving voxelization (IFPV) in simplifying the representation of the dose distribution. Methods: The concept of IFPV was introduced for concise representation of a treatment plan. IFPV is a sparse voxelization scheme that partitions the voxels into subgroups according to their geometric, anatomical, and dosimetric features. In this study a deep CNN was constructed to buildup a dose prediction model in IFPV domain based on 60 volumetric modulated arc therapy (VMAT) treatment plans from a database of previously treated 70 prostate cancer patients. The dose prediction model learns the contour to dose relationship and predicts the dose distribution in IFPV domain given the input contours. Additional ten independent prostate cases were selected as testing data.DVH comparison, dose difference maps, and residual analysis with the sum of absolute residual (SAR) were used to evaluate the performance of the proposed method. Results: The proposed IFPV-based method achieved good prediction performance in terms of DVH comparison and dose difference maps. Statistical results of SARs showed that the IFPV-based method is comparable with voxel-based method even though the number of dose representation points used in the IFPV-based method was substantially reduced. The proposed approach achieved mean SARs of 0.029 AE 0.020 and 0.077 AE 0.030 for bladder and rectum, respectively, compared with mean SARs of 0.039 AE 0.029 and 0.069 AE 0.028 in the conventional voxel-based method. Conclusions: A novel deep CNN-based dose prediction method in IFPV domain was proposed. The proposed approach has great potential to significantly improve the efficiency of dose prediction and facilitate the treatment planning workflow.

Dose Prediction Using a Three-Dimensional Convolutional Neural Network for Nasopharyngeal Carcinoma With Tomotherapy

“…Some deep-learning-based dose prediction studies have been made for cervical carcinoma. The studies used a general 3D-model-patch-training strategy with 16 pixels height matrix to train (shape of n×n×16) or directly used a 2D network for data training ( 5 , 31 , 32 ). From some dose prediction studies proposed, 2D network training is good enough to provide excellent results of dose prediction.…”

The Application and Development of Deep Learning in Radiotherapy: A Systematic Review

“… 55 Similar experience in predicting dose distribution for prostate cancer was validated in a plan for 80 prostate cancer patients. 56 Although the current clinically available ML-based automatic plan effectively saved time, the generated plan still had to be corrected manually. In the future, DL-based AP commercial software is expected to generate plans that can directly meet the needs of clinical treatment.…”

Probabilistic feature extraction, dose statistic prediction and dose mimicking for automated radiation therapy treatment planning

“…The first part of predicting dose-related quantities has been abundantly addressed in previous literature. Regarding spatial dose prediction, two-or three-dimensional convolutional neural networks in various architectures, such as U-nets, have been widely used [9][10][11][12][13] and extended to generative models such as generative adversarial networks [14,15], although other methods such as random forests have also been studied [16]. For prediction of DVHs or other dose statistics, while overlap volume histograms evaluated on the input image have been traditionally used for this purpose [17][18][19][20][21][22][23][24], more recent literature also includes the use of neural networkbased methods to simultaneously predict spatial dose and DVHs directly from input images [25,26].…”