New brain tumor AI pipeline matches top algorithms with modular design

A team of researchers led by Daniel Capellán-Martín and Marius George Linguraru has introduced a new adaptable segmentation pipeline for brain tumors that rivals top algorithms without being tied to a specific neural network architecture. Presented at MICCAI BraTS 2025, the pipeline tackles the challenge of segmenting diverse adult and pediatric brain tumors from multi-parametric MRI, including data from the BraTS 2025 Lighthouse Challenge: multi-consortium pediatric tumors (PED), preoperative meningiomas (MEN), meningioma radiotherapy (MEN-RT), and pre-/post-treatment brain metastases (MET).

The key innovation is a modular workflow that selects and combines state-of-the-art models, then applies tumor- and lesion-specific processing before and after training. Radiomic features extracted from MRI scans help identify tumor subtypes, ensuring more balanced training across heterogeneous datasets. Custom lesion-level performance metrics determine each model's influence in the ensemble and guide post-processing refinements tailored to each case. On the BraTS test sets, the pipeline achieved results comparable to top-ranked algorithms, confirming that lesion-aware processing and flexible model selection can deliver robust segmentations—potentially enabling quantitative tumor measurement in clinical practice for diagnosis and prognosis.