ESTRO 2024 - Abstract Book

S4432

Physics - Machine learning models and clinical applications

ESTRO 2024

ensures treatment quality and safety, benefiting from ongoing learning and improvement through the integration of additional patient data into the system.

Material/Methods:

The study focuses on developing and validating three treatment models:

i. Prostate Model: Concentrating on prostate cancer treatment with dose levels between 66 Gy and 78 Gy in 200 cGy fractions. The model, trained with data from 73 VMAT-treated patients using Varian's Eclipse software, accounts for high and low dose PTVs, bladder, rectum, and femoral heads. Validation involved re-planning 26 patients with RapidPlan, emphasizing parameters like maximum PTV dose, bladder and rectum doses, and femoral head maximum doses. ii. Single Isocenter Radiosurgery (SRS) Model for Multiple Cranial Metastases: Designed for treating multiple brain metastases with a single isocenter, utilizing data from 31 VMAT-treated patients with Eclipse. The model encompasses PTVsoma, healthy brain tissue, cochlea, optic pathways, brainstem, and optimization rings.[3] Validation included 24 RapidPlan-replanned patients, with assessments covering Compliance Index, Paddick Compliance Index, Gradient Index, and volumes receiving 10 Gy and 12 Gy in healthy brain tissue.[4][5][6][7] iii. Lung Radiosurgery Model (SBRT) for Peripheral Lesions: Addressing peripheral lung metastases with SBRT, trained with data from 25 VMAT-treated patients using Varian's Eclipse and 4D CT images. The model considers the target lesion, lungs, bronchus, trachea, great vessels, heart, ribs, and optimization rings. Validation involved re-planning 16 patients with RapidPlan, assessing Compliance Index, High Dose Spillage, Gradient Index, and Low Dose Spillage.[8] Consistent prescription criteria were maintained across all models. Statistical analyses, including normality tests and one-tailed Student's t-tests (see Tables 1 and 2), were conducted for each model. These models and rigorous analyses aim to enhance the precision and safety of radiation therapy across various cancer treatments. In prostate cases, we compared the original planning to Radiplan. Dose reductions was found, such as Dmax for the bladder, D15%, D25%, D35%, and D50% for both the bladder and rectum, with p-values ranging from 0.001 to 0.050. Additionally, Dmax for the right and left femoral heads decreased, with p-values of 0.013 and 0.017, while D2% in the original plan remained unchanged (p-value 0.38). These results demonstrate an enhanced overall dose distribution with Radiplan. For cases of multiple brain metastases, the RapidPlan model significantly reduced radiation doses to critical structures, including healthy brain tissue, optic pathway, cochleas, and brainstem, with supporting p-values ranging from 0.022 to 0.05. Plan parameters, like IC RTOG and IC Paddick, improved, with p-values of 0.049 and 0.004, respectively. However, the Gradient Index evaluation showed no substantial change, with a p-value of 0.143. In summary, Radiplan also improved dose distribution. In the lung model analysis, we assessed Compliance Index, High Dose Spillage, Gradient Index (Low Dose Spillage), and Low Dose Spillage (2cm from the target), with p-values ranging from 0.28 to 0.42. Although no significant improvements were observed, the model's results closely mirrored those of the original plans, confirming its validity. Results: