Problem

The integration of artificial intelligence (AI) and radiomics into clinical practice for brain tumor research is hampered by fragmented workflows, a lack of transparency, and insufficient alignment with end-user needs. Despite advancements in methodological aspects like model architecture and training strategies, operational transferability and seamless integration into clinical workflows remain underdeveloped. This often leads to promising experimental solutions staying confined to laboratory prototypes, with minimal impact on real-world clinical decision-making and patient care.

Method

The authors developed the initial version of a scalable, web-based visual analytics system to facilitate radiomics-driven machine learning inference specifically for neuro-oncology. This platform consolidates three primary functions into a unified interface: managing patient cohorts from structured clinical data, extracting radiomic features from medical images and segmentation masks, and conducting guarded inference with pre-trained machine learning models. The system was designed through an iterative, user-centered process, emphasizing lightness and reproducibility, and was implemented using the Dash library. A key design principle involved the explicit exposure of intermediate workflow artifacts, thereby enhancing traceability, interpretability, and responsible AI utilization.

Results

The system, designated V0, underwent evaluation using both the public BraTS 2020 glioblastoma dataset and a proprietary clinical cohort of skull base meningioma patients. It comprises three interconnected modules: the Data module, which supports cohort ingestion, interactive exploration, and extension; the Radiomics module, enabling the upload of MRI volumes, segmentation masks, and PyRadiomics configuration files for feature extraction and seamless merging with clinical data; and the Prediction module, which provides on-demand inference using embedded models. This module includes explicit guardrails to prevent the application of incompatible models, displaying warnings to users. Overall, V0 establishes an end-to-end operational layer that links cohort handling, radiomics extraction, and safeguarded prediction within a single dashboard, with each component designed to ensure inspectability of intermediate artifacts and auditable actions.

Implications

This proposed framework offers a practical foundation for clinically oriented AI applications in brain tumor analysis by enhancing transparency, traceability, and the responsible use of AI. Its implementation as a lightweight web application aligns with the realistic constraints of clinical environments, promoting portability and easy integration even with limited hardware resources. However, the system currently faces limitations, including its constraint to specific data formats (CSV/XLSX for tabular data and NIfTI for imaging/masks), the potential for slow online radiomic feature extraction on non-dedicated hardware, and incomplete functional coverage in its initial version. Future work will focus on extending compatibility to more common clinical formats like DICOM, optimizing feature extraction performance, and continuous refinement based on clinical feedback to improve applicability and functional completeness.