ESTRO meets Asia 2024 - Abstract Book

S401

RTT – Treatment planning, OAR and target definitions

ESTRO meets Asia 2024

Keywords: AI, DLC, lymph region

References:

* Mirada Medical Ltd., Oxford, United Kingdom

** MVison AI, Helsinki, Finland

*** Limbus AI, Canada

317

Digital Poster

Multicenter evaluation of deep-learning based deliverable automated radiotherapy planning

Lei Yu 1,2 , Jiazhou Wang 1,2 , Weigang Hu 1,2

1 Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China. 2 Department of Oncology, Shanghai Medical College, Fudan University, Shanghai, China

Purpose/Objective:

Automated treatment planning (ATP) based on deep learning (DL) dose prediction holds great promise, but still is rarely accessible in clinical scenarios, because ATP cannot always perfectly satisfy the personalized requirements from physicians, even though volumetric dose distribution of the patient is accurately predicted. In this study, we developed a hybrid ATP strategy by combining DL dose prediction and physicians’ requirements to generate directly applicable plans of various cancer sites in a commercial TPS.

Material/Methods:

We employed a channel attention densely-connected U-Net (CAD-UNet) architecture [1] to predict patient-specific dose distribution. DVH indices were extracted from the predicted dose by referring to a given goal sheet that represents the physician’s concerns (i.e., trade-offs between target coverage and OAR sparing with priorities, same for each site), and used as optimization objectives to generate an executable automated plan. DL models for common tumor sites including breast, cervix, rectum, head&neck, and lung were trained and validated based on the historical cases of our institution (institution A). The models and the goal sheets were implemented in a commercial TPS and tested for clinical use in other two institutions (B and C). Minor modifications of the goal sheets might be needed to meet the respective criteria of the institution. The generated ATP plans were reviewed in the subjective and objective criteria, with the enrollment of 30, 10, and 10 patients per site for the institution A, B and C, respectively.

Results:

A complete set of data from all the institutions is still under processing. The preliminary assessment results of the institution A demonstrate that above 70% of the automated plans passed the clinical criteria after one round of optimization (<5 min), and in the blinded review among three expert physicians, more than a half of the automated