| Author: | Xiong, Tianyu |
| Title: | Automatic treatment planning for functional lung avoidance radiotherapy |
| Advisors: | Cai, Jing (HTI) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Health Technology and Informatics |
| Pages: | xxi, 154 pages : color illustrations |
| Language: | English |
| Abstract: | Background: Functional Lung Avoidance Radiotherapy (FLART) is a promising radiotherapy (RT) technique that may reduce pulmonary toxicity compared to conventional lung RT. Unlike conventional anatomy-guided RT, FLART features in utilizing functional images to account for lung function heterogeneity and preferentially sparing high-function lung (HFL) regions for better lung function protection. The incorporation of lung function information into treatment planning is crucial for FLART. However, the majority of current FLART studies rely on manual planning. The significant inter-patient variability and intra-lung heterogeneity of lung function distributions, coupled with limited clinical experience as an emerging RT technique, render manual FLART planning highly challenging. Therefore, manual FLART planning is not only time-consuming and labor-intensive but also prone to producing inconsistent or suboptimal FLART plans. Purpose: This study aims to develop a novel multi-modality-guided dose prediction (MMDP)-based automatic planning (auto-planning) algorithm for FLART and evaluate its clinical performance and potential benefits. Method: The study is structured in three parts. First, we developed a fully automatic planning algorithm for FLART. It integrates a dosimetric score-based beam angle selection method and a meta-optimization (MO)-based plan hyperparameter optimization method, both of which incorporate lung function information to guide dose redirection from HFL to low-functional lung (LFL). It is applicable to both contour-based FLART (cFLART) and voxel-based FLART (vFLART) optimization options. A cohort of 18 lung cancer patients underwent planning CT and SPECT perfusion (Q) scans were collected to preliminarily evaluate the performance of the MO-based auto-planning algorithm. In the second part, we further developed an MMDP-based auto-planning algorithm for FLART, building on the MO-based method. We collected data from 196 lung cancer patients across three institutions, comprising 114/28 cases with lung ventilation (V) surrogate maps for training/validation and 21/33 cases with SPECT V/Q images for testing. The MO-based auto-planning algorithm was employed to generate ConvRT (MO-ConvRT) and FLART (MO-FLART) plans. High-quality plans were selected to train a novel MMDP model which features in extracting complementary features from multi-modality inputs for predicting optimal dose distributions. An innovative instance-weighting anatomy-to-function training strategy was tailored to enhance prediction accuracy. A function-guided dose mimicking algorithm was developed to convert predicted dose distributions into FLART (MMDP-FLART) plans, which were compared against MO-based and manual plans. In the third part, the proposed auto-planning algorithms were implemented in clinical treatment planning systems (TPS) and their clinical performance and benefits were evaluated. An additional 83 pairs of high-quality manual ConvRT/FLART plans were created for fine-tuning (50 pairs) and evaluation (33 pairs) of the auto-planning algorithms. Clinical performance was assessed using quantitative DVH/DFH metrics, normal tissue complication probabilities, and qualitative blind review by one senior medical physicist and two senior radiation oncologists. Additionally, the FLART planning results by a junior planner with and without the assistance of the developed MMDP-FLART system were compared, aiming at evaluating the potential of the system in facilitating the broader clinical adoption of the emerging FLART technique. Results: In the first part, automatic ConvRT plans generated by the MO-based approach exhibited similar quality compared to manual counterparts. Furthermore, compared to automatic ConvRT plans, HFL mean dose, V20, and V5 were significantly reduced by 1.13 Gy (p<.001), 2.01% (p<.001), and 6.66% (p<.001) respectively for cFLART plans. The vFLART plans showed a decrease in functionally weighted mean lung dose (fMLD) by 0.64 Gy (p<.01), F20 by 0.90% (p=0.099), and F5 by 5.07% (p<.01) respectively. Though inferior conformity was observed, all dose constraints were well satisfied. The ablation study results indicated that both function-guided beam angle selection and plan optimization significantly contributed to dose redirection. In the second part, the MMDP model achieved accurate dose predictions, with DVH score of 1.94 Gy, Dose score of 1.14 Gy, and prediction errors for fMLD of 0.10±0.57 Gy. In contrast, MMDP trained through naïve training and an anatomy-guided dose prediction model significantly overestimated (p<.001) fMLD by 0.40±0.58 Gy and 0.30±0.60 Gy respectively, validating the effectiveness of the proposed training strategy and multi-modality learning. Compared to manual FLART plans, MMDP-FLART plans exhibited significantly lower (p<.001) and comparable (p=0.099) fMLD on SPECT V and Q datasets respectively. Compared to manual ConvRT plans, MMDP-FLART plans effectively achieved lung dose painting and reduced fMLD by 0.80 Gy (11.9%, p<.001) and 0.46 Gy (6.0%, p<.001) on SPECT V and Q datasets respectively. In the third part, both MO-based and MMDP-based auto-planning algorithms were integrated into clinical TPS. Compared to manual ConvRT plans, MMDP FLART plans reduced fMLD and HFL Dmean by 11.8% and 15.1%, respectively. For FLART-benefiting patients, MMDP FLART plans reduced the probability of grade ≥2 radiation pneumonitis by 6.25 percentage points (27%). The MMDP-based algorithm demonstrated superior lung function protection and clinical acceptability compared to the MO-based approach. Blind review by three senior clinicians revealed that 88% of MMDP FLART plans were clinically acceptable without modification, and 69% were rated as comparable or superior to manual FLART plans created by an eight-year-experience senior planner. In terms of planning efficiency, manual FLART planning required approximately 120–180 minutes per plan, whereas MO-based and MMDP-based auto-planning required approximately 78 minutes and 8 minutes per plan, respectively. Furthermore, assistance from the auto-planning system substantially enhanced the quality of FLART plans generated by a junior planner (with one year of planning experience), elevating the plan quality to a level comparable to that of a senior planner (with eight years of planning experience). Conclusion: An innovative auto-planning algorithm for FLART based on multi-modality-guided dose prediction has been technically developed, clinically deployed, and comprehensively evaluated. It shows considerable promise in enhancing the efficiency, consistency, and quality of FLART planning, while also reducing radiation-induced lung toxicity compared to the current clinical standard. A demo for the clinically ready auto-planning system is available at the link. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14711