3D radiotherapy dose prediction on head and neck cancer patients with a hierarchically densely connected U-net deep learning architecture

Abstract

The treatment planning process for patients with head and neck (H&N) cancer is regarded as one of the most complicated due to large target volume, multiple prescription dose levels, and many radiation-sensitive critical structures near the target. Treatment planning for this site requires a high level of human expertise and a tremendous amount of effort to produce personalized high quality plans, taking as long as a week, which deteriorates the chances of tumor control and patient survival. To solve this problem, we propose to investigate a deep learning-based dose prediction model, Hierarchically Densely Connected U-net, based on two highly popular network architectures: U-net and DenseNet. We find that this new architecture is able to accurately and efficiently predict the dose distribution, outperforming the other two models, the Standard U-net and DenseNet, in homogeneity, dose conformity, and dose coverage on the test data. Averaging across all organs at risk, our proposed model is capable of predicting the organ-at-risk max dose within 6.3% and mean dose within 5.1% of the prescription dose on the test data. The other models, the Standard U-net and DenseNet, performed worse, having an averaged organ-at-risk max dose prediction error of 8.2% and 9.3%, respectively, and averaged mean dose prediction error of 6.4% and 6.8%, respectively. In addition, our proposed model used 12 times less trainable parameters than the Standard U-net, and predicted the patient dose 4 times faster than DenseNet.reduce the vanishing gradient issue, and decrease the number of trainable parameters needed. While the term "densely connected" was historically used to described fully connected neural network layers, this publication by Huang et al. had adopted this terminology to describe how his convolutional layers were connected. While requiring more memory to use, the authors showed that the DenseNet was capable of achieving a better performance while having far less parameters in the neural network. For example, they were able to have comparable accuracy with ResNet, which had 10 million parameters, using their DenseNet, which had 0.8M parameters. This indicates that DenseNet is far more efficient in feature calculation than existing network architectures. For its contribution to the AI community, the DenseNet publication was awarded for the CVPR 2017 best publication. However, it is recognized that DenseNet, while efficient in parameter usage, actually utilizes considerably more GPU RAM, rendering a 3D U-net with fully densely connected convolutional connections infeasible for today's current GPU technologies.Motivated by a 3D densely connected U-net, but requiring less memory usage, we developed a neural network architecture that combines the essence of these two influential neural network architectures into our proposed network while maintaining a respectable RAM usage, which we call Hierarchically Densely Connected U-net (HD U-net). The term "hierarchically" is used here to describe the different levels of resolution in the U-n...

Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose‐volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy

“…These were chosen by evaluating the order of magnitude of the values that each loss function exhibits for a converged model. From previous dose prediction studies and results, 48,55 we can estimate that the L MSE $ 10 À4 and L DVH $ 10 À3 for a converged model. Since we are using least squares GAN framework, we estimate the loss L ADV G ranges from 10 À1 to 10 0 .…”

“…With the advancements in deep learning, particularly in computer vision [44][45][46] and convolutional neural networks, 47 many studies have investigated clinical dose distribution prediction using deep learning on several sites such as for prostate IMRT, 48,49 prostate VMAT, 32 lung IMRT, 50 head-andneck IMRT, [51][52][53][54] and head-and-neck VMAT. 55 In addition to clinical dose prediction, deep learning models are capable of accurately generating Pareto optimal dose distributions, navigating the various tradeoffs between planning target volume (PTV) dose coverage and organs at risk (OAR) dose sparing. 56 Most of these methods utilize a simple loss function for training the neural networkthe mean squared error (MSE) loss.…”

Technical Note: A feasibility study on deep learning‐based radiotherapy dose calculation

“…The general HD U-net structure was proposed for three-dimensional dose prediction for head and neck cancer patients. 9 We adopted the main neural network operations of the original HD U-net and made modifications to the architecture. Patient's CT image and corresponding RT dose distribution are used as model inputs and the CS dose distribution is the output.…”

“…7 The model was also modified to predict the dose distributions for head and neck cancer patients and lung cancer patients with heterogeneous beam setups. [7][8][9] In this work, we explore the feasibility of using DL for accurate and fast radiotherapy dose calculation. Specifically, we test the Hierarchically Densely Connected U-net (HD U-net) model for intensity-modulated radiation therapy (IMRT) dose calculation for prostate cancer patients using precalculated low-accuracy dose distributions as the model input.…”

Knowledge‐based automated planning with three‐dimensional generative adversarial networks

“…8,17,18 The most recent work in this space has focused on neural network-based KBP methods, which are trained on libraries of historical plans to predict dose for each axial slice separately [i.e., two-dimensional (2D) KBP methods] 8,18,19 or all slices concurrently [i.e., three-dimensional (3D) KBP methods]. 20,21 Among the 2D methods, generative adversarial networks (GANs) have been shown to perform the best 18 while among the 3D methods, DoseNet is considered state-of-the-art. 20 It remains an open question as to whether a combination of the two approaches, a 3D GAN, will achieve even better results.…”