Experiences on using the CHAIMELEON secure processing environment in an open competition addressing five AI...

The CHAIMELEON Repository is a secure and interoperable platform designed to accelerate the adoption of Artificial Intelligence (AI) in cancer imaging research by enabling the sharing and harmonisation of large-scale medical, clinical, and molecular datasets across Europe. The objective is to evaluate how effectively the repository supports the development of predictive AI models for five major cancer types - prostate, lung, breast, colon, and rectum - through an Open Challenge. Participants developed models using platform-provided data, with performance assessed on key clinical tasks such as risk stratification, survival prediction, and tumour staging.

The Open Challenge was a two-phase international competition. Participants developed predictive AI models using computational resources, imaging and clinical data provided through the repository. Models were evaluated on risk classification, survival prediction, and tumour staging, within a GDPR-compliant, cloud-based environment integrating model development, evaluation, and resource monitoring.

The Open Challenge engaged 482 registered participants from 29 countries, of which 162 downloaded the datasets and 42 submitted validation and/or testing processes during the Classification Phase. During the Championship Phase, participants executed 3591 batch jobs consuming more than 2433 computational hours across GPU- and CPU-based queues. The repository supported AI model development and evaluation for five cancer-related tasks, including risk stratification, survival prediction, TNM staging, and histological classification. The highest test scores achieved during the Classification Phase were 0.6894 for prostate cancer, using a composite score combining AUC, sensitivity, specificity, and balanced accuracy, and 0.7699 for lung cancer, using the Concordance Index (C-index).

The CHAIMELEON Open Challenge demonstrated the feasibility of a secure, scalable, and GDPR-compliant infrastructure for collaborative AI development in oncology. The repository successfully supported heterogeneous clinical AI tasks across five cancer types while enabling standardised benchmarking, reproducible experimentation, and controlled access to harmonised multimodal datasets and computational resources. The observed participant activity and computational workload further confirmed the platform's capacity to operate under realistic multi-user research conditions and highlighted its potential as a infrastructure for future large-scale cancer imaging studies.