Three-Dimensional Radiotherapy Dose Prediction on Head and Neck Cancer Patients with a Hierarchically Densely Connected U-net Deep Learning Architecture (1805.10397v3)

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.

Three-Dimensional Radiotherapy Dose Prediction with Hierarchically Densely Connected U-net

The paper presented focuses on a significant challenge in the field of head and neck (H&N) cancer treatment, particularly radiotherapy dose prediction. The proposed method harnesses deep learning advancements to design an efficient model that reduces the time and expertise required for planning complex treatments. This approach leverages the Hierarchically Densely Connected U-net (HD U-net), drawing from the architectures of both the U-net and DenseNet.

Overview and Methodology

Treatment planning for H&N cancer is intricate, attributed to the large target volumes, diverse prescription levels, and numerous critical organs at risk (OARs). Current planning processes, involving substantial human intervention and lengthy optimization cycles, can inadvertently compromise tumor control due to delays. The paper seeks to alleviate this bottleneck by utilizing the HD U-net, a deep learning model incorporating hierarchical and dense connections to predict three-dimensional dose distributions efficiently and accurately.

The HD U-net architecture combines the U-net's capability of capturing global and local features with the DenseNet's efficient feature reuse and propagation. The model is constructed to reduce the computational load associated with 3D predictions, making the model feasible on current GPU technologies. The training incorporated data from 120 patients with various H&N cancer sites using detailed patient-specific data without normalizing prescription doses for the input channels, ensuring the model learns to handle the variability in dose prescriptions autonomously.

Results

The HD U-net model performs exceptionally compared to its precursors, the Standard U-net and DenseNet. Notably, it achieves superior accuracy with significantly fewer trainable parameters. On average, HD U-net predicts the maximum dose to OARs within 6.3% and the mean dose within 5.1% of the prescribed dose, outperforming the other models, which have larger prediction errors. HD U-net's design allows for this increased accuracy using 12 times fewer trainable parameters than the Standard U-net and predicting doses four times faster than the DenseNet.

The model's efficacy is further highlighted by its performance metrics—demonstrating superior homogeneity, dose conformity, and dose coverage. It maintains a moderate computational expense, striking a balance between prediction accuracy and processing time.

Implications and Future Directions

The implications of this paper are profound in both practical and theoretical domains of radiotherapy. Practically, employing the HD U-net for dose prediction could significantly streamline planning workflows, reducing treatment delays and possibly improving patient outcomes. Theoretically, the model's design principles can inspire further innovations in architecture design—melding different model strengths to manage complexity and efficiency better.

Future research directions will likely explore incorporating CT imaging and additional clinical constraints into the model to improve predictions further. Additionally, the integration of this approach into a fully AI-based treatment planning system appears a promising development avenue. Such advancements could potentially surpass traditional methods in both efficiency and clinical outcomes, potentially revolutionizing the therapeutic planning landscape in radiology.

In conclusion, the HD U-net presents a robust solution to the existing challenges in H&N radiotherapy planning, demonstrating the potential of deep learning to enhance medical treatment processes. Continued exploration and refinement of this approach could lead to significant enhancements in personalized cancer therapy planning and delivery.