3445 - Feasibility of an AI-Driven Auto-Contouring and Auto-Planning Workflow for Lung Cancer Radiotherapy with Radiation Oncologist Quality Assurance
Presenter(s)
C. Choi1, N. P. Mankuzhy2, J. Willmann3, G. Jhanwar1, S. Elguindi1, M. Zarepisheh1, J. Jiang1, M. Thor1, and H. Veeraraghavan1; 1Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, 2Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 3Department of Radiation Oncology, University Hospital Zurich, University of Zurich, Zurich, Switzerland
Purpose/Objective(s): The purpose of this feasibility study was to evaluate an end-to-end automated contouring and radiotherapy (RT) planning framework and to determine whether treatment plans derived from auto-contoured tumors achieve comparable plan quality based on physician edited tumors.
Materials/Methods:
Ten patients with lung tumors who previously underwent intensity-modulated RT were included. The plan CT and the corresponding structure set were collected. The gross tumor volume (GTV) was auto-segmented on the plan CT using a published AI model and edited by a radiation oncologist. Automated replanning was performed for both the AI-automated (AI-auto) and edited contours (AI-assisted) using an in-house hierarchical optimization-based algorithm (ECHO), applying the same planning objectives and optimization parameters as the clinically delivered plans. AI and edited GTVs were geometrically evaluated against clinical contours (i.e. original contours used in RT) using dice similarity coefficient (DSC) and total added path length (APL). Dose to the planning target volume (PTV) (created by adding 12mm isotropic margin to AI- automated and AI-assisted GTV), heart, esophagus, and lungs were compared across all three plans using the Wilcoxon signed-rank test.Results:
Automated GTV segmentation failed in two of ten patients: one due to misdetection, and one due to misidentification of pleural fluid as tumor. These cases were excluded from the comparisons. The DSC accuracy improved from 0.83 ± 0.10 with AI-automated to 0.87 ± 0.06 AI-assisted contouring. For GTVs ranging between 13 - 311 cc, the APL for editing AI GTVs ranged between 82cm to 491cm. As shown in Table I, there was no difference in dose metrics for the PTV using either AI-automated or AI-assisted plans as well as for the esophagus. On the other hand, the mean dose increased to the heart (AI-auto: 8.9, AI-assisted: 9.1 vs. original: 7.6 Gy, p<0.05) and the lung (AI-assisted: 12 vs. original: 11 Gy, p<0.05). Further analysis of the case with the largest dose deviation showed high DSCs (AI-auto: 0.95 vs. AI-assisted: 0.94); however, higher doses in the AI-auto and AI-assisted plans (heart: 14.3 and 14.9 Gy; lungs: 17.5 and 17.1 Gy) compared with the original plan (heart: 8.6 Gy; lungs: 15.6 Gy) were attributed to smaller PTV margins in the original treatment, which resulted in lower heart and lung doses. Table 1: Dosimetric comparison between the AI-auto, AI-assisted replans and original ECHO plans.Conclusion: We demonstrated that AI-automated segmentation methods can support automated replanning; however, clinical expertise remains essential to ensure appropriate PTV margin selection and treatment adaptation for effective implementation.
| Original ECHO plan | AI-auto ECHO replan | AI-assisted ECHO replan | |
| D95% PTV (Gy) | 60 ± 2.8 | 60 ± 2.8 (p=0.06) | 60 ± 2.8 (p=0.06) |
| Dmean Heart (Gy) | 7.6 ± 7.2 | 8.9 ± 8.0 (p=0.02*) | 9.1 ± 8.5 (p=0.04*) |
| Dmean Esophagus (Gy) | 6.8 ± 6.2 | 7.1 ± 5.7 (p=0.95) | 7.7 ± 6.2 (p=0.31) |
| Dmean Lungs (Gy) | 11 ± 3.2 | 12 ± 3.4 (p=0.06) | 12 ± 3.6 (p=0.02*) |