ESTRO 2024 - Abstract Book

S4446

Physics - Machine learning models and clinical applications

ESTRO 2024

[2] H. Chen, C. Gomez, C.-M. Huang, and M. Unberath, “Explainable medical imaging AI needs human-centered design: guidelines and evidence from a systematic review,” Npj Digit. Med., vol. 5, no. 1, Art. no. 1, Oct. 2022, doi: 10.1038/s41746-022-00699-2.

[3] M. J. Page et al., “The PRISMA 2020 statement: an updated guideline for reporting systematic reviews,” Syst. Rev., vol. 10, no. 1, p. 89, Mar. 2021, doi: 10.1186/s13643-021-01626-4.

[4] T. Speith, “A Review of Taxonomies of Explainable Artificial Intelligence (XAI) Methods,” in 2022 ACM Conference on Fairness, Accountability, and Transparency, Seoul Republic of Korea: ACM, Jun. 2022, pp. 2239–2250. doi: 10.1145/3531146.3534639.

[5] T. Miller, “Explanation in artificial intelligence: Insights from the social sciences,” Artif. Intell., vol. 267, pp. 1–38, Feb. 2019, doi: 10.1016/j.artint.2018.07.007.

886

Digital Poster

Enabling Tradeoff Preference in Deep Learning (DL) based HN IMRT Auto Planning

Q. Jackie wu, xinyi li, Yang Sheng, qiuwen wu

Duke University, Radiation Oncology, Durham, USA

Purpose/Objective:

Our previously developed AI planning tool demonstrated promising plan quality, robustness, and efficiency for patients for planning HN cases with bilateral primary PTVs. Each AI prediction takes a few seconds, and total planning time from contours to plans takes about 10 min (mostly in data transfer). However, the prototype model only generates one plan for each patient, and thus, the dosimetric tradeoff is fixed in the AI plan. This project aims to incorporate tradeoff options in these AI models.

Material/Methods:

Two approaches are developed to enable dose tradeoff in AI planning. The first one is model set approach. A series of AI models were trained with several groups of well-orchestrated ground truth plans, where each group of plans represents a typical dose tradeoff in clinical scenarios. More specifically, the most common dose tradeoffs for HN were generalized into two classes, one is between PTV coverage (labeled “P”) and OAR sparing (labeled “O”). The other one is between the left and right parotids (“L”, “R”, and “B” for sparing both parotids with same priority). More scenarios can be considered and modeled in future by expanding to more tradeoff classes. The other approach is called modeling approach, where OAR priorities and patient anatomy are both used as AI model inputs, so that AI models can generate plans with different dose tradeoffs for the same patient. During the training phase, AI plans with randomized priorities were generated and used as ground truth. The AI model thus learns the relationship of PTV coverage and OAR sparing priorities and the corresponding dose distribution/fluence map