2344 - Control Dosimetric Outcome of AI-Generated Plans through OAR Prioritization for Head-and-Neck Cancer: An Investigation on Input Sensitivity

X. Li¹, D. Yang¹, Y. Sheng¹, Y. Ge², Q. Wu¹, and Q. J. J. Wu³; ¹Duke University Medical Center, Durham, NC, ²The University of North Carolina at Charlotte, Charlotte, NC, ³Duke University, Durham, NC

Purpose/Objective(s): Examine response of an Artificial Intelligence (AI) model for Head-and-Neck (HN) IMRT treatment planning to various organs-at-risk (OAR) priorities as model inputs. Materials/

Methods: Our lab has developed an AI system that predict IMRT plan’s optimal fluence maps based on patient’s anatomical information and user-defines OAR priorities. This AI system was specialized for HN primary targets with bilateral segments and showed promising dosimetric outcomes using preset OAR priorities as inputs. This study focuses on examining the response of the AI system to an extensive range of OAR priorities as inputs to investigate if the model prediction breaks under certain conditions and if the AI behaves as expected in the extrapolation region. During training of the AI system, 2 IMRT plans were generated for each of the 200 training cases. One used a set of fixed priorities that have balanced dose tradeoffs among OARs and PTV; the other one used the same set of objectives whereas priorities were randomly altered by up to To examine the changes in response, the priorities of 4 OARs including L&R parotid, oral cavity, and cord+5mm were altered as inputs. Each priority was altered by -80%, -40%, 0%, 40%, and +80% from the balanced values, and a total of 5^4=625 AI plans were generated for a test case. The AI plans had full dose calculated by a treatment planning system. All plans were normalized as 100% prescription dose covers 90% of PTV. To avoid extreme cases, the AI plans with conformity index > 1.1, heterogeneity index > 30, or max dose > 125% were excluded from statistics calculations. Median dose of L&R parotid, median dose of oral cavity, and D1cc of cord+5mm was evaluated. These metrics were linearly fit using OAR priorities as inputs and 5-fold cross validation.
Results: 232 OAR priority sets were included in the statistics. These sets do not include any -80% priority change, indicating the AI model breaks when users heavily prioritize PTV over OARs. Among these 232 AI plans, L&R parotids median dose was altered by almost ±10Gy. Ground truth plan with balanced tradeoffs was used as reference. Linear fitting coefficients averaged among the 5 folds were reported in the table in the format of y = c1 + c2 x ?P(parotidR) + c3 x ?P(parotidL) + c4 x ?P(oralcavity) + c5 x ?P(cord5mm), where ?P(OAR) was the ratio of priority variation. R-squared of linear fitting was ~0.95 except for cord+5mm.
Conclusion: The AI system could generate substantially different AI plans for the same patient and the response was essentially linear to the input priorities. This study demonstrated the promising performance of the AI system and potentially enhanced trust in AI’s clinical applications. Table 1

	Reference	Min	25%	50%	75%	Max	c1	c2	c3	c4	c5	r2
Parotid R Dmedian (Gy)	16.2	11.3	13.9	18.7	25.6	30.3	21.4	-12.7	-2.2	1.0	0.1	0.95
Parotid L Dmedian (Gy)	15.0	13.2	15.5	19.7	24.0	30.0	21.8	-1.9	-10.9	-1.4	0.5	0.96
Oral Cavity Dmedian (Gy)	22.4	20.8	23.3	27.1	29.8	32.9	27.8	-0.3	0.6	-7.8	-0.5	0.94
Cord+5mm D1cc (Gy)	24.9	22.8	24.1	25.2	26.5	28.8	25.4	0.3	1.0	0.4	-2.6	0.88