Non‐convex IMRT treatment planning using deep inverse optimization

Background Intensity‐modulated radiation therapy (IMRT) utilizes inverse optimization to calculate beamlet intensities, achieving clinical dose objectives for tumors while minimizing dose to organs‐at‐risk (OARs). However, incorporating non‐convex dose‐volume constraints into IMRT planning is challe...

Full description

Saved in:

Bibliographic Details
Published in:	Medical physics (Lancaster) Vol. 52; no. 11; pp. e70141 - n/a
Main Authors:	Lei, Yang, Zhang, Jiahan, Liu, Tian
Format:	Journal Article
Language:	English
Published:	United States 01.11.2025
Subjects:	Carcinoma, Non-Small-Cell Lung - radiotherapy deep inverse optimization Deep Learning Humans IMRT auto‐planning Lung Neoplasms - radiotherapy non‐convex programming Organs at Risk - radiation effects Radiotherapy Dosage Radiotherapy Planning, Computer-Assisted - methods Radiotherapy, Intensity-Modulated - methods IMRT auto‐planning non‐convex programming deep inverse optimization
ISSN:	0094-2405, 2473-4209, 2473-4209
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Background Intensity‐modulated radiation therapy (IMRT) utilizes inverse optimization to calculate beamlet intensities, achieving clinical dose objectives for tumors while minimizing dose to organs‐at‐risk (OARs). However, incorporating non‐convex dose‐volume constraints into IMRT planning is challenging due to their combinatorial nature and computational complexity. Purpose This study explores a novel approach to IMRT fluence map optimization (FMO) by leveraging deep inverse optimization (DIO). Our method aims to approximate non‐convex constraints efficiently while preserving convexity, improving adherence to clinical dose objectives. Methods We introduce a new relaxation technique that preserves convexity by incorporating second‐order cone constraints to approximate non‐convex conditions. A DIO framework is then employed to solve the second‐order cone programming problem, generating clinically feasible treatment plans. To balance tumor coverage and OAR sparing, we employ sequential optimization to integrate multiple objectives, ensuring effective dose distributions across patient datasets. The method was evaluated retrospectively on 30 locally advanced non‐small cell lung cancer (NSCLC) patients treated with 60 Gy in 30 fractions. Plan quality was assessed using dosimetric metrics, including D98%, D2%, and maximum dose for the planning target volume (PTV), as well as dose‐volume percentages, mean, and maximum doses for the lung, heart, esophagus, and spinal cord. Additionally, optimization convergence complexity and duty cycle were analyzed to assess planning and delivery efficiency. Results Compared to traditional convex optimization, the proposed method maintains comparable PTV coverage while significantly improving dose homogeneity (narrower D98%‐to‐D2% range, reduced hotspots). It enhances OAR sparing, reducing lung V5Gy (p = 0.01), heart V30Gy (p = 0.04), and spinal cord maximum dose (p < 0.01), while ensuring clinically acceptable dose distributions. Additionally, the optimization process demonstrates significantly faster convergence (∼5 vs. ∼39 iterations). Conclusion This study presents an effective and computationally efficient approach to IMRT treatment planning by approximating non‐convex dose‐volume constraints with second‐order cone programming and DIO. The method enhances plan quality by improving dose homogeneity and OAR sparing while significantly accelerating optimization convergence, offering a practical solution for advanced radiation therapy planning.
Bibliography:	ObjectType-Article-1 SourceType-Scholarly Journals-1 ObjectType-Feature-2 content type line 23
ISSN:	0094-2405 2473-4209 2473-4209
DOI:	10.1002/mp.70141