Swin UNETR++: Advancing Transformer-Based Dense Dose Prediction Towards Fully Automated Radiation Oncology Treatments (2311.06572v3)

Abstract: The field of Radiation Oncology is uniquely positioned to benefit from the use of artificial intelligence to fully automate the creation of radiation treatment plans for cancer therapy. This time-consuming and specialized task combines patient imaging with organ and tumor segmentation to generate a 3D radiation dose distribution to meet clinical treatment goals, similar to voxel-level dense prediction. In this work, we propose Swin UNETR++, that contains a lightweight 3D Dual Cross-Attention (DCA) module to capture the intra and inter-volume relationships of each patient's unique anatomy, which fully convolutional neural networks lack. Our model was trained, validated, and tested on the Open Knowledge-Based Planning dataset. In addition to metrics of Dose Score $\overline{S_{\text{Dose}}}$ and DVH Score $\overline{S_{\text{DVH}}}$ that quantitatively measure the difference between the predicted and ground-truth 3D radiation dose distribution, we propose the qualitative metrics of average volume-wise acceptance rate $\overline{R_{\text{VA}}}$ and average patient-wise clinical acceptance rate $\overline{R_{\text{PA}}}$ to assess the clinical reliability of the predictions. Swin UNETR++ demonstrates near-state-of-the-art performance on validation and test dataset (validation: $\overline{S_{\text{DVH}}}$=1.492 Gy, $\overline{S_{\text{Dose}}}$=2.649 Gy, $\overline{R_{\text{VA}}}$=88.58%, $\overline{R_{\text{PA}}}$=100.0%; test: $\overline{S_{\text{DVH}}}$=1.634 Gy, $\overline{S_{\text{Dose}}}$=2.757 Gy, $\overline{R_{\text{VA}}}$=90.50%, $\overline{R_{\text{PA}}}$=98.0%), establishing a basis for future studies to translate 3D dose predictions into a deliverable treatment plan, facilitating full automation.

• The paper introduces Swin UNETR++ with a lightweight 3D Dual Cross-Attention module that enhances dense dose prediction in radiation oncology.
• The model achieved quantitative DVH and Dose Scores of approximately 1.492 Gy and 2.649 Gy on validation, with high clinical acceptance rates on test data.
• It integrates patient imaging and segmentation for voxel-level prediction, supporting the automation of complex radiation treatment plan generation.

The paper entitled "Swin UNETR++: Advancing Transformer-Based Dense Dose Prediction Towards Fully Automated Radiation Oncology Treatments" addresses a critical challenge in the field of Radiation Oncology—automating the generation of radiation treatment plans for cancer therapy. This task involves integrating patient imaging with organ and tumor segmentation to develop a 3D radiation dose distribution that aligns with clinical treatment goals, which is akin to voxel-level dense prediction.

The authors introduce Swin UNETR++, an advanced model based on the transformer architecture that incorporates a lightweight 3D Dual Cross-Attention (DCA) module. This module is designed to capture both intra- and inter-volume relationships specific to each patient's anatomy, a capability that traditional fully convolutional neural networks generally lack.

The model was extensively trained, validated, and tested using the Open Knowledge-Based Planning dataset. The performance of Swin UNETR++ was assessed using both quantitative and qualitative metrics:

1. Quantitative Metrics:
   • Dose Score ($\overline{S_{\text{Dose}}$) and DVH Score ($\overline{S_{\text{DVH}}$): These metrics measure the difference between the predicted and the actual (ground-truth) 3D radiation dose distributions.
   • On the validation dataset, the model achieved:
     • $\overline{S_{\text{DVH}} = 1.492 \, \text{Gy}$
     • $\overline{S_{\text{Dose}} = 2.649 \, \text{Gy}$
   • On the test dataset, it obtained:
     • $\overline{S_{\text{DVH}} = 1.634 \, \text{Gy}$
     • $\overline{S_{\text{Dose}} = 2.757 \, \text{Gy}$
2. Qualitative Metrics:
   • The paper introduces two new qualitative metrics to evaluate the clinical reliability of the model:
     • Average Volume-Wise Acceptance Rate ($\overline{R_{\text{VA}}$)
     • Average Patient-Wise Clinical Acceptance Rate ($\overline{R_{\text{PA}}$)
   • In the validation set, the model demonstrated:
     • $\overline{R_{\text{VA}} = 88.58\%$
     • $\overline{R_{\text{PA}} = 100.0\%$
   • In the test set, these rates were:
     • $\overline{R_{\text{VA}} = 90.50\%$
     • $\overline{R_{\text{PA}} = 98.0\%$

These results highlight Swin UNETR++ as nearly state-of-the-art in automating radiation dose prediction, underscoring its potential to facilitate full automation of radiation treatment plan creation. The combination of high quantitative performance and robust clinical acceptance rates provides a solid foundation for future research aimed at translating 3D dose predictions into practical, deliverable treatment plans.