QP 23 - CNS 3: Advances in Glioma, Spine, and Normal Tissue Toxicity
1131 - Macropathology and Multimodal Magnetic Resonance Image Based Deep Learning Model for Personalized Definition of Target Volume in Patients with Glioma
Shandong Cancer Hospital and Institute Jinan, Shandong
N. Shulun1,2, J. Xu2, Y. Su3, L. Wang3, and M. Hu2; 1Cheeloo College of Medicine, Shandong University, Jinan, Shandong, China, 2Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China, 3Department of Radiation Oncology Physics and Technology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China
Purpose/Objective(s): The common methods employed to define the clinical target volume in gliomas primarily rely on conventional magnetic resonance (MR) imaging. However, these approaches have notable limitations as they predominantly focus on the contrast-enhancing component of the tumor, often neglecting numerous microscopic extensions. Artificial intelligence (AI) has emerged as a promising tool to aid in precise radiotherapy planning. The aim of this study was to develop a deep learning model utilizing macropathology and multimodal MR imaging to enhance personalized target delineation for glioma.Materials/
Methods: Twenty-three patients diagnosed with gliomas (5 diffuse astrocytomas, 10 oligodendrogliomas, and 8 glioblastomas) undergoing supra-total resection were prospectively enrolled. At initial diagnosis, all patients underwent multimodal MR images, which included T1-weighted, contrast-enhanced T1-weighted, T2-weighted, T2 fluid-attenuated inversion recovery, diffusion-weighted imaging, and MR spectroscopy. Firstly, H&E sections were acquired from the surgical specimen, and regions-of-interest such as the tumor parenchyma, infiltrating regions, central necrotic areas, normal brain parenchyma, and peritumoral edema regions were delineated. These regions were subsequently registered to multimodal MR images and parcellated into patches. Features were then extracted from these patches utilizing the ResNet-50 deep learning framework, and a glioma invasive model with AI (GLIM) was developed. To assess its clinical applicability further, we retrospectively gathered data from 117 glioma patients to serve as a real-world cohort. We compared GLIM model-derived tumor volume and recurrence coverage with standard radiotherapy plans. Model performance was evaluated through area under the curve (AUC), sensitivity, specificity, Dice, and Jaccard coefficient. Results: A total of 810,000 patches were obtained and randomly divided into training (567,000 patches) and validation (243,000 patches) groups at a ratio of 7:3. The GLIM model achieved robust performance, with an AUC of 0.881 (95% CI, 0.877 to 0.884) for tumor prediction in the training cohort, surpassing radiologist performance by 29.8% in sensitivity and 4.8% in specificity. The GLIMs favorable performance was confirmed in the validation cohorts. Particularly noteworthy is the significant improvement observed in recurrence coverage within the real-world cohort, achieved without a notable increase in radiation volume, by utilizing GLIM-derived target volumes instead of standard radiation plans, with a mean Jaccard index of 86.2% and a mean Dice score of 88.7%. Conclusion: The GLIM highlights the potential of tumor invasive modeling for individualized therapy of glioma, which could enhance the accuracy of radiation planning in the future.
