S2318
Physics - Machine learning and AI algorithms
ESTRO 2026
techniques: IMRT, hybrid IMRT and hybrid VMAT. Radiat Oncol 17, 60 (2022). https://doi.org/10.1186/s13014-022-02009-2Karaca S. The use of hybrid techniques in whole-breast radiotherapy: A systematic review. Technol Cancer Res
Treat. 2022;21:153303382211439. doi:10.1177/153303382211439. Keywords: Breast, Prediction, Index
Results: Med-ImageNet currently supports queries across all TCIA collections with associated metadata and establishes explicit links between paired imaging modalities (e.g., CT with RTSTRUCTs). Users can query and request datasets based on imaging region and imaging modality, and download appropriate datasets using their TCIA credentials. Med-ImageNet processes the raw DICOM files to generate AI-ready imaging outputs, tabular files with image-specific metadata, as well as a dataset summary. These outputs are broadly applicable for training, evaluating, and benchmarking machine learning models across diverse imaging research tasks. We have used the AI-ready images in two models: (1) an auto-segmentation model for head and neck cancer using nnU-Net architecture (Isensee et al., 2020) which is now deployed in clinical silent mode and (2) to conduct bias analyses of the foundation model MedSAM2 (Ma et al., 2025), (3) and to create quality control pipelines for radiomic feature extraction methods such as PyRadiomics (van Griethysen et al., 2017). Conclusion: Med-ImageNet transforms heterogeneous cancer image collections into a harmonized, AI-ready repository for oncology research, while enabling integration of user-provided imaging data. By combining reproducible indexing, standardized preprocessing, and open-source tooling, Med- ImageNet addresses a critical bottleneck in cancer imaging AI research and provides a scalable foundation for developing fairer, more generalizable oncology machine learning models. The generated outputs have already supported diverse applications, from training segmentation models to bias auditing and radiomics quality control, underscoring their broad utility across research domains. References: Clark, K et al. (2013). The Cancer Imaging Archive (TCIA). https://doi.org/10.1007/s10278-013-9622-7De Biase et al. (2024). Clinical. https://doi.org/10.1093/bjrai/ubae015Haibe-Kains et al. (2020). Transparency. https://doi.org/10.1038/s41586-020-2766-yIsensee, F et al. (2020). nnU-Net.https://doi.org/10.1038/s41592- 020-01008-zKim S et al. Med-ImageTools. (2025). https://doi.org/10.12688/f1000research.127142.3Koça k, B et al. (2019). Radiomics with artificial intelligence.
Digital Poster Highlight 4846 Med-ImageNet – A Standardized Resource for AI- Ready Oncology Imaging Joshua Siraj 1 , Jermiah Joseph 1 , Sejin Kim 1 , Declan Korda 1 , Muammar Kabir 1 , Katy Scott 1 , Mattea Welch 1 , Tran Truong 1 , Amber Simpson 2 , Tony Tadic 1 , Andrew Hope 1 , Benjamin Haibe-Kains 1 , Clare McElcheran 1 1 Princess Margaret Cancer Center, University Health Network, Toronto, Canada. 2 Biomedical and Molecular Sciences, Queens University, Kingston, Canada Purpose/Objective: Artificial intelligence (AI) and radiomics features have demonstrated potential for advancing oncology through improved diagnosis, prognosis, and treatment personalization (Koçak et al., 2019). However, heterogeneous cancer imaging datasets, characterized by inconsistent preprocessing, variable metadata, and limited reproducibility, remain a barrier to model development (Haibe-Kains et al., 2020) and clinical translation (De Biase et al., 2024). We aimed to develop an open-source platform that transforms heterogeneous collections into harmonized, AI-ready resources tailored for oncology research. Material/Methods: We developed Med-ImageNet, an open-source platform that harmonizes publicly available TCIA collections (Clark et al., 2013) and user-provided datasets into an AI-ready repository. The platform comprises of three components: (1) Med-ImageDB for dataset indexing, query API, and secure images and metadata retrieval; (2) Med-ImageTools (Kim et al., 2025) for standardized preprocessing including DICOM ingestion, voxel harmonization, intensity normalization, and metadata alignment; and (3) the Med-ImageNet repository unifying these modules into a scalable and reproducible data compendium. The system supports both raw data access and AI-ready outputs (e.g., NIfTI format) for deep learning integration.
Made with FlippingBook - Share PDF online